DEV Community: Pineprompt

Copilot rank tracking: what to track instead

Pineprompt — Fri, 16 Oct 2026 15:54:28 +0000

Copilot rank tracking is a slightly misleading phrase. A Copilot answer is written text, not a list of ten blue links, so there is no position one to hold. What you can track is whether Copilot mentions your brand, where in the answer it appears, which sources it cites, and how it describes you. Measure those four things across a fixed set of prompts, repeated over time, and you have the Copilot equivalent of a rank tracker.

A Copilot visibility tracker automates this. This guide explains what to measure so you can judge any tool, or do it by hand.

Why "rank" does not carry over

In Bing results, a URL has a position. Position 3 is position 3 until something changes. You can check it daily and compare.

A Copilot answer works differently.

It is generated, not retrieved. Copilot writes a response for each prompt. Microsoft's documentation says that when Copilot uses web search, it parses the prompt, identifies terms where a search would improve the response, and generates a search query that it sends to the Bing Search service. The answer is then built from what comes back and from the model.

It varies. Ask the same question twice and the wording, the brands named and their order can differ.

The unit is a brand, not a page. Buyers read "consider A, B or C." They do not click through ten results. Your goal is to be in that sentence.

It depends on the prompt. "Best CRM for startups" and "which CRM suits a five-person team" are different prompts. A single keyword-style rank cannot describe both.

So replace "what is our rank for this keyword" with "how often, where and how are we included across these prompts."

The four things to track instead

1. Mention

Does the answer name your brand at all? This is the base signal. Record yes or no per run, then calculate a mention rate: runs with a mention divided by total runs.

Do it per prompt and across the full set. A 60% mention rate on category prompts and 5% on comparison prompts tells you where to work.

2. Position within the answer

If Copilot lists several brands, order matters. First mention gets more attention than fifth. Record where you appear: first, middle, last, or in a closing "also consider" line.

This is the nearest thing to rank. Treat it as a rough signal, not a precise number. Because answers vary, average position over many runs rather than trusting one.

3. Citations

Copilot answers can link to their sources. Microsoft's privacy documentation notes that stored interaction data includes Copilot's response along with citations to information used to ground it. In practice, check each answer for its source links.

Log two things:

Your own pages cited. Which URLs, on which prompts.
Other domains cited. Review sites, forums, news, competitor blogs.

The second list is often the more useful. It shows where Copilot looks when it answers questions about your category, and so where you need to be present.

4. Sentiment and framing

Being named is not enough. Read how you are described. Is the brand recommended, listed neutrally, or mentioned with a caveat? Tag each mention as positive, neutral or negative, and note recurring phrases such as "expensive" or "best for small teams."

How to track Copilot by hand

Start with a spreadsheet.

Write 20 to 50 prompts. Cover category, use-case, comparison and problem questions. Use real buyer wording. Keep the set fixed.
Run each prompt three or more times in fresh conversations. Use the same surface each time, whether that is Copilot on the web, in Edge or in Windows.
Log each run. Date, prompt, run number, mentioned yes or no, position, competitors named, cited URLs, sentiment.
Calculate weekly. Mention rate, average position, citation rate, top cited domains, sentiment split.
Change one thing, then rerun. Publish a page, fix a listing, earn a mention. Compare the next week to the baseline.

The same process works for other engines. Our guide to tracking your brand in ChatGPT walks through it in more detail.

Do Bing rankings still matter for Copilot?

Somewhat. Because Copilot sends web search queries to Bing when it searches, being findable in Bing is a sensible foundation. If your pages are not indexed there, they are less likely to be available as sources.

But a good Bing ranking does not guarantee a Copilot mention. The answer is assembled from several sources and shaped by the prompt. Keep tracking Bing rankings if you already do. Just do not treat them as a stand-in for answer visibility. Measure the answers directly.

What a good Copilot tracker should do

If you are comparing tools, check for these:

Repeated runs per prompt, not one-off checks.
Mention, position, citation and sentiment as separate fields.
Cited URLs and domains, not only a score.
Competitor tracking across the same prompts.
History, so you can see trends and the effect of changes.
Other engines in the same view, since buyers use several.

Be cautious about any tool that reduces all of this to a single "Copilot rank." It hides the variation that makes the data meaningful.

A worked example of the log

Say you run one prompt, "best project management tool for a small agency," three times. In run one, Copilot names four tools and you are second. In run two, you are not named. In run three, you are named last, with a link to a third-party review that cites you.

Your log for that prompt reads: mentioned in two of three runs, a 67% mention rate. Positions were second and fourth. Cited domains: one review site, none of your own. Sentiment: neutral both times.

That is a useful result. It says you are in the conversation but not reliably, and the source Copilot leans on is a review page you do not control. The next step is clear: get that review updated, or earn coverage on similar pages.

Common mistakes

Trusting one run. One answer proves nothing. Use repeats.

Changing prompts mid-stream. You lose comparability.

Ignoring wording. Track the phrases Copilot uses about you, not just presence.

Skipping citations. Without source URLs you cannot tell what to fix.

Tracking only your brand. Without competitors, a mention rate has no context.

Where Pineprompt fits

Pineprompt runs a fixed prompt set against Copilot on a schedule and records mentions, position, citations and sentiment, with your competitors alongside. Pineprompt tracks six live AI platforms daily: ChatGPT, Gemini, Claude, Google AI Overviews, Google AI Mode and Copilot, and all six are on every plan. Plans start at $99 a month; see pricing. If a spreadsheet does the job for now, use that. Either way, track answers, not ranks.

Originally published on pineprompt.com

How to track brand mentions in Perplexity

Pineprompt — Fri, 16 Oct 2026 15:48:47 +0000

To track mentions in Perplexity, write a fixed set of the questions your buyers ask, run each one several times, and log two things per run: whether your brand is named in the answer, and which source URLs Perplexity cites. Repeat weekly. That is the whole method. The rest of this guide is how to do it without fooling yourself.

If you would rather not do it by hand, a Perplexity visibility tracker runs the same process on a schedule.

Why Perplexity is the easiest engine to audit

Perplexity's own help centre says each response includes citations and links to original sources. That matters for tracking. You get two signals from one answer:

A mention. Your brand name appears in the answer text.
A citation. One of your pages appears in the source list.

They are not the same. A brand can be named without any of its pages being cited. A page can be cited without the brand being named. Track both.

The citations also answer a second question people ask: how to see which sources Perplexity is drawing on. Open the answer, read the source list, and write the URLs down. Those are the pages shaping what buyers read.

Step 1: Build a prompt set

Write 20 to 50 prompts. Use the words buyers use, not your marketing copy. Cover four types:

Category prompts. "Best [category] tools."
Use-case prompts. "[Category] for [audience or job]."
Comparison prompts. "[Your brand] vs [competitor]."
Problem prompts. "How do I [problem you solve]?"

Sales calls, support tickets and community threads are the best source of wording. If you want to widen the set, the prompt fan-out generator turns one seed question into the variations buyers might type.

Keep the set fixed once you start. If you change the prompts every week, you cannot compare weeks.

Step 2: Run every prompt more than once

Perplexity answers can change between runs. The sources it picks can change too. One run is an anecdote.

Run each prompt at least three times, in fresh threads, so earlier answers do not colour later ones. Use the same account state and the same mode each time. Note which mode you used. If you switch between modes partway through, your numbers stop being comparable.

Step 3: Log what you see

Use a spreadsheet. One row per run. Columns:

Column	What to record
Date	When you ran it
Prompt	Exact wording
Run number	1, 2, 3
Brand mentioned	Yes or no
Position	Order in the answer, if it lists brands
Competitors named	Which ones
Cited URLs	Every source URL shown
Your URLs cited	Any of your own pages
Framing	Recommended, neutral, or criticised

Copy the URLs in full. You will want to sort them by domain later.

Step 4: Work out what to measure

Raw logs are noise until you turn them into a few numbers.

Mention rate. Runs where your brand appears, divided by total runs. Calculate it per prompt and across the whole set. This is your headline number.

Citation rate. Runs where at least one of your own pages is cited, divided by total runs. If mention rate is high and citation rate is low, Perplexity knows you exist but is reading other people's pages about you.

Share of voice. Your mentions compared with each competitor's, across the same runs. It shows who owns the category in Perplexity's answers.

Top cited domains. Group all cited URLs by domain and count them. Review sites, forums, news pages and competitor blogs will show up. This list tells you where to earn coverage. Our guide on how to get your brand cited by Perplexity covers what to do with it.

Framing. Read how you are described. A mention that calls you outdated is not a win.

Trend. Compare week to week. A single week is a snapshot. Four weeks is a direction.

What to do with the numbers

Look for three patterns:

Absent on important prompts. You are missing from category prompts where competitors appear. Check which pages Perplexity cites for those prompts and ask what they have that you do not.
Mentioned, never cited. Your own pages are not being used. Check that they answer the question plainly and that the crawler can reach them.
Cited, but poorly framed. A third-party page describes you badly. Fix the source if you can.

Change one thing at a time. Then rerun the same prompt set. Otherwise you will not know what worked.

Where manual tracking breaks down

The spreadsheet method is honest and cheap. It also has limits.

Time. 30 prompts, three runs each, is 90 runs a week for one engine. Add competitors and other engines and it becomes a job.

Consistency. People skip weeks, reword prompts and forget to log. The data gets patchy.

Scale. Logging URLs and counting domains by hand gets slow fast.

Coverage. Buyers do not use only one engine. If you want the same view for ChatGPT, Gemini, Claude and others, you repeat the whole exercise for each. Our guide to tracking your brand in ChatGPT shows the same method for a different engine.

A reasonable rule: track by hand until the process is stable and you know which prompts matter. Automate when the weekly chore is the reason you stop doing it.

Common mistakes

Trusting one run. A single answer is a sample of one. Repeat every prompt.

Rewording prompts between weeks. You lose the ability to compare. Fix the set, and add new prompts as a separate group.

Counting only mentions. A mention with no citation and a mention with three citations to your own pages are very different outcomes.

Ignoring competitors. A 40% mention rate means little until you know a competitor sits at 70%. Log who else is named on every run.

Forgetting the source pages. The cited URLs are the most actionable data you will collect. Keep them, and review them monthly.

A simple starting plan

Write 20 prompts this week.
Run each three times and log the results.
Calculate mention rate, citation rate and the top cited domains.
Pick the weakest prompt group and make one change.
Rerun in a week and compare.

That gives you a baseline, and a baseline is what most teams lack.

Where Pineprompt fits

Pineprompt runs prompt sets like this on a schedule and records mentions, position, sentiment and cited sources. Pineprompt tracks six live AI platforms daily: ChatGPT, Gemini, Claude, Google AI Overviews, Google AI Mode and Copilot, and all six are on every plan. Plans start at $99 a month; see pricing. If you prefer to keep the spreadsheet going, the method above works fine. You can also start with the Perplexity visibility tracker page to see what automated tracking looks like.

Originally published on pineprompt.com

How to track your brand in Google AI Mode

Pineprompt — Fri, 16 Oct 2026 14:21:31 +0000

Google AI Mode is a separate tab in Google Search that answers questions in a conversation. It writes a long answer, links sources, and lets the user ask follow-ups. To track your brand in it, choose a fixed set of prompts, run them on a schedule, and record whether you are named, where, how you are described and which pages are linked.

This guide explains how AI Mode builds its answers, how to check it by hand, what to measure and how to improve your results.

How AI Mode sources its answers

AI Mode uses a Gemini model on top of Google Search. Its key technique is query fan-out. Instead of running your question as one search, it breaks it into several related sub-queries and runs them at the same time. It then reads the results and combines them into one response. Some sub-queries look at a topic from different angles. Others check specific facts, such as prices or reviews.

An example. A user asks, "What is the best CRM for a five person consultancy?" AI Mode might search for CRM tools for small teams, CRM pricing, CRM reviews from consultants, and integrations with common email tools. Your brand has to show up in the results of at least some of those searches to be considered.

This is close to prompt fan-out in other assistants, but AI Mode draws on the Google index directly, so classic search signals carry more weight than in most other engines. For a general definition, see Google AI Mode. It also differs from an AI Overview, which is a short summary on the normal results page. AI Mode is longer, conversational and more likely to compare options.

The practical result: you do not compete for one query. You compete across a cluster of them. Pages that rank for the sub-queries, and state their facts clearly, are the ones that get pulled in.

Choose the prompts to track

Use 20 to 50 prompts written as real users would type them. AI Mode invites longer questions than standard search, so include those.

Cover four types:

Category: "Best accounting software for freelancers in Germany"
Comparison: "Xero vs QuickBooks for a small agency"
Situation: "I run a five person team and need to track client work. What should I use?"
Follow-up: a second question that builds on the first answer

Follow-ups matter. AI Mode keeps context, so a brand that is missing from the first answer may appear in the second. Record the full sequence, and keep the wording fixed between runs.

Split branded and non-branded prompts in your log. Non-branded prompts show discovery. Branded prompts show what the engine says about you once someone asks.

Check AI Mode manually

Open Google in a private window and select the AI Mode tab. Note your country and language. Results vary by place.
Enter one prompt. Use a fresh session for each prompt.
Record the date, the location, the prompt and the full answer.
Note whether your brand is named, its position among the options, and the wording used about you.
Record every linked source: the URL, the domain, and whether it is yours, a competitor's, or a third party.
If you can see the searches behind the answer, note them. They show the fan-out, and they tell you which queries you need to win.
Repeat the prompt three times.

Then ask the follow-up and repeat the steps.

Repeat runs are not optional. Answers change between runs and between days. See answer volatility. A single run tells you very little.

What to measure

Mentions. How often your brand is named across runs. This is your brand mention rate.

Position. Where you sit in the list of options. AI Mode often presents ranked or grouped picks. Being first or in the opening group counts for more. See position in answer.

Sentiment. The tone, and the accuracy, of what is written about you. Log wrong facts separately, since a brand hallucination needs a specific fix.

Citations. Which pages AI Mode links. Split them into your own pages, competitor pages and third party pages. Keep mentions and citations as separate columns. A brand can be named without a link, and linked without being recommended. The citation rate shows how often your pages are used.

Share of voice. Your mentions as a share of all brand mentions across the prompt set. It shows how you compare with the competitors you named in advance. See share of voice in AI.

For AI Mode, add one more log: the fan-out queries, where visible. Count how many of them you rank for in normal search. That gap often explains a low mention rate.

Where manual tracking breaks down

Manual checks work for a first baseline. They struggle with scale. Twenty prompts, three runs each, plus follow-ups, is well over a hundred answers per round. Location adds more. Copying links and reading long answers takes hours, and the data ends up inconsistent.

Pineprompt tracks six live AI platforms daily: ChatGPT, Gemini, Claude, Google AI Overviews, Google AI Mode and Copilot. It stores each answer with mentions, position, sentiment and citations, so you can see trends and compare engines. The Google AI Mode tracking page has the details. There is no free trial. See pricing for plans.

How to improve your visibility in AI Mode

Because AI Mode leans on Google Search, much of the work is search work aimed at a wider set of questions.

Map the fan-out for your topics. For each key prompt, list the sub-queries likely behind it: the category, the comparison, the price, the reviews, the integrations, the alternatives. Check whether you have a page that answers each one.

Cover the sub-questions on your own site. Build focused pages, one question each, with the answer at the top. A short pricing page, a clear comparison, a use case page for each audience. Pages that state facts plainly are easier to use in a combined answer.

Keep pages indexable. Make sure Googlebot can crawl and index them, and that they render without needing a script to show the main text. Check your robots.txt and canonical tags.

Earn third party coverage. Many fan-out queries return review sites, forums and listings, not brand pages. Look at the cited domains in your log. If a site appears repeatedly and you are absent or described badly, that is a target.

Use consistent facts and structure. Keep your name, category, prices and features identical across your site and profiles. Add structured data where it fits. This helps Google resolve you as one entity.

Show experience. Original data, named authors, real customer examples and dates help. See E-E-A-T in AI search.

Keep content current. AI Mode fetches live results, so a fresh, updated page can be picked up within weeks.

A simple weekly routine

Run the prompt set and follow-ups, three runs each.
Update the five measures and the fan-out notes.
Pick the three prompts with the largest gap to a competitor.
For each, find the missing sub-query page or listing, and ship it.
Review the trend after four weeks.

Log each change with its date, so you can link it to what happens next.

Summary

AI Mode turns one question into many searches and writes one answer from the results. Track it with a fixed prompt set, follow-ups and repeated runs. Measure mentions, position, sentiment, citations and share of voice, and note the fan-out queries. Improve by covering the sub-questions, keeping pages indexable and getting listed on the sources it cites. Begin by hand, then move to a tool once the volume is too much.

Originally published on pineprompt.com

How to track Google AI Overviews

Pineprompt — Fri, 16 Oct 2026 14:05:17 +0000

An AI Overview is the generated summary Google shows above the normal results for some searches. To track it, you pick a fixed set of queries, check each one on a schedule, and record whether an Overview appears, whether your brand is named, which sources are linked and how you are described. Repeat over time and you can see a trend.

That is simple to say and tedious to do. This guide covers the manual method, what to measure, when a tool makes sense, and what no method can promise. If you want the tool route now, see our page on Google AI Overview tracking.

Why classic rank tracking falls short

A rank tracker tells you the position of a URL in the results. An AI Overview is not a position. It is written text, often with a few source links beside it.

Three things follow from that.

Not every query gets one. Google shows Overviews for some queries and not others. Whether one appears can change for the same query from week to week.

The answer varies. The wording, the brands named and the links shown can differ between runs.

Ranking well does not guarantee inclusion. A page can rank in the normal results and never be named in the Overview. The reverse can also happen.

So the questions change. Not "what is our rank," but "when an Overview appears for this query, are we in it, and how?"

What to measure

Record the same five things for every query, every time.

1. Overview present

Did an AI Overview appear at all? Record yes or no. Your trigger rate is the share of checks where one appeared. If a query never triggers an Overview, it is not an Overview problem.

2. Brand mention

Is your brand named in the Overview text? Your mention rate is the runs with a mention divided by the runs where an Overview appeared. Keep this separate from the trigger rate, or you will mix up "no Overview" with "Overview without us."

3. Cited sources

Overviews link to pages they draw on. Note two lists:

Your pages cited. Which URLs, for which queries.
Other domains cited. Publishers, review sites, forums, competitors.

The second list is often the more useful one. It shows who Google leans on for your topic, and so where you may need to be present.

4. Position and framing

If the Overview names several brands, note where you sit: first, middle, last. Then read how you are described. Tag each mention as positive, neutral or negative, and write down recurring phrases such as "budget option" or "best for small teams."

5. Competitors

Log which competitors appear in the same Overview. Share of voice is your mentions as a share of all brand mentions across the set. It says more than a raw count.

The manual method

You can do this with a spreadsheet and an hour a week. Here is a workable process.

Build a query list. Start with 20 to 30 queries a buyer would type. Mix category queries ("best project management software"), comparisons ("tool A vs tool B") and problem queries ("how to stop missing deadlines"). Our prompt fan-out generator can help you widen the list.
Fix the conditions. Use a private window, and stay signed out. Use the same country and language for every check. Results differ by location.
Run each query and record the five measures. Save a screenshot. You will want it later.
Repeat on a schedule. Weekly is the minimum that shows a trend. Answers move, so a single check proves little.
Summarise. Calculate trigger rate, mention rate, and the top cited domains.

A row in your log might look like this:

Query	Overview	Named	Position	Sentiment	Cited
best crm for startups	Yes	Yes	2nd of 4	Neutral	Two review sites, one competitor
crm pricing comparison	No	n/a	n/a	n/a	n/a

Where the manual method breaks

It works for a small set. It stops working when:

The list grows. Thirty queries once a week is 120 checks a month. Add countries or languages and it multiplies.
Results differ by place. Checking from one office tells you about one location.
You need volume for confidence. One run is an anecdote. Ten runs of the same query start to show a pattern.
Someone else needs the numbers. Screenshots in a folder are hard to turn into a report.
Consistency slips. Checks done by different people, on different days, under different conditions do not compare cleanly.

When a tool makes sense

Use a tool when the checking is the bottleneck. A good one should:

Run your query set on a fixed schedule, daily if you need it.
Target a country and language, so you see what your buyers see.
Record Overview presence, mentions, position, sentiment and cited sources for each run.
Show competitors and share of voice next to your own numbers.
Keep history, so you can compare this month with last.
Cover the other engines too. Buyers do not stick to one.

Pineprompt tracks prompts daily across ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, Google AI Mode, Grok and Copilot. It reports mentions, position in the answer, sentiment, cited sources, competitors and share of voice, with country and language targeting. Our methodology page explains how the checks are run and recorded. Plans start at $99 a month, and there is no free trial; see pricing.

Do not forget AI Mode

AI Overviews are one Google surface. AI Mode is another: a conversational search experience where a follow-up question continues the thread. It handles longer, more specific questions, and it works differently from an Overview. Track it separately. Our page on Google AI Mode tracking covers it.

Both surfaces also break a question into several smaller searches behind the scenes. That is called prompt fan-out, and it affects which pages get pulled into the answer.

The limitations

Be honest about what tracking can and cannot tell you.

It is a sample, not a census. You see the answers to your queries, not every query your buyers use. Choose them with care and revisit the list.

Answers vary. Two runs can differ. That is why repeated runs and averages beat single checks.

Personalisation and location matter. A signed-in user in another city may see something different from your log. Treat your numbers as a consistent yardstick, not a mirror of every user's screen.

No tool can show you Google's internal reasoning. You can see the output and the links. You cannot see why a page was chosen. Any claim that it can is worth doubting.

Tracking does not change the result. It tells you where you stand. Improving it is separate work: clear pages, sources that back you up, and presence on the sites Google cites.

Common mistakes

Changing the query list each week. Keep it fixed so the trend means something. Add new queries as a separate group.
Counting "no Overview" as a loss. If none appears, there is nothing to be in.
Checking once. One run is not a result.
Ignoring the cited domains. They tell you where to act.
Only tracking your brand name. Buyers ask about categories and problems, not just you.

A simple starting plan

This week, write 25 queries and run each one by hand. Record the five measures. Next week, repeat. After a month, you will know your trigger rate, your mention rate and the three domains Google cites most. That is enough to decide whether to keep going by hand or move to a tool. Either way, start with a fixed list and measure the same way each time.

Originally published on pineprompt.com.

How to track your brand in Google Gemini

Pineprompt — Fri, 16 Oct 2026 13:56:18 +0000

Gemini is Google's AI assistant. People use it in the Gemini app, in Chrome, on Android and inside Google Workspace. When they ask it for a tool, a vendor or a comparison, it names brands. To track yours, pick a fixed set of prompts, run them on a schedule, and record whether Gemini mentions you, where, how it describes you and which sources it cites.

This guide covers how Gemini builds an answer, how to check it by hand, what to measure and what you can do to improve the result.

How Gemini sources its answers

Gemini is a large language model. It learned from a large body of text up to a cutoff date. That gives it general knowledge about your brand if you were well covered before then. It also gives it a knowledge cutoff, so anything newer is missing unless it looks it up.

For many questions Gemini does look things up. It can use Google Search to fetch current pages and then write an answer based on them. Google calls this grounding. The model sends search queries, reads the results, and cites the pages it used. This is a form of retrieval-augmented generation.

Two things follow from this.

First, Gemini has two routes to your brand. One is memory from training. The other is live retrieval from the Google index. A brand can be strong on one and weak on the other. A new product may be invisible in memory but show up through grounding. An old brand may be remembered with outdated facts.

Second, Gemini does not always ground. Simple or well known questions may be answered from memory with no sources. Questions about prices, recent news, or "best" lists are more likely to trigger a search. You can read more about the difference in training data vs live retrieval.

Because the Google index feeds grounding, ordinary search work still matters. Pages that are crawlable, indexed and clear about what they say are the ones Gemini can use.

Choose the prompts to track

Start with 20 to 50 prompts. Write them the way a buyer would ask, not the way a marketer would.

Use a mix:

Category prompts: "What are the best project management tools for small agencies?"
Comparison prompts: "Compare Tool A and Tool B for a team of ten."
Problem prompts: "How do I reduce churn in a subscription app?"
Branded prompts: "What is Pineprompt and who is it for?"

Keep branded and non-branded prompts separate in your records. Branded prompts show what Gemini believes about you. Non-branded prompts show whether you are found by people who do not know you yet.

Freeze the wording. If you change a prompt, you lose the trend. Add new prompts as new rows, not edits.

Check Gemini manually

A manual check is enough to start.

Open Gemini in a private window, signed out if you can. Signed-in history and personalisation can shape answers.
Paste one prompt. Start a new chat for each prompt so earlier turns do not leak in.
Record the date, the model version shown in the app, and the prompt.
Copy the full answer into a sheet.
Note whether your brand appears, at what position in any list, and how it is described.
Open the source links, if any, and record the domains.
Run the same prompt three times. Answers vary, so one run is an anecdote.

Do this for each prompt, then repeat weekly. It takes time. Twenty prompts at three runs each is sixty answers, and that is one snapshot.

Also check the free app against the paid tiers if your buyers use both, and check the different model options. They can behave differently. Record which one you used every time.

What to measure

Five measures cover most of what you need.

Mentions. The share of runs in which your brand is named at all. Call it brand mention rate. It is the base number. If it is zero, nothing else matters yet.

Position. Where you appear in a list or recommendation. First place carries more weight than fifth. See position in answer.

Sentiment. Whether the wording is positive, neutral or negative, and whether it is accurate. Record wrong claims too. A confident false statement is a brand hallucination and it needs a fix.

Citations. Which URLs Gemini links when it grounds. Track your own pages and third party pages separately. Note the difference between being mentioned and being cited. A mention with no link still shapes opinion. A citation sends traffic and shows which of your pages the model trusts.

Share of voice. Your mentions divided by all brand mentions across your prompt set, compared with named competitors. It puts your number in context. See share of voice in AI.

Report each as an average across repeated runs, not a single result. Expect answer volatility. A move of a few points week to week is usually noise. A steady change over a month is a signal.

Where manual tracking breaks down

The method works, but it does not scale. Sixty answers a week is already an hour or more of copying and reading. Add competitors, more prompts, other regions and other engines, and it stops being practical. Sheets also lose detail: the source list, the exact wording, the model version.

Most brands also need to look beyond Gemini. Buyers use several assistants, and they disagree with each other. Pineprompt tracks six live AI platforms daily: ChatGPT, Gemini, Claude, Google AI Overviews, Google AI Mode and Copilot, and stores the answers, mentions, position, sentiment and citations for each run. The Gemini visibility tracker page shows what that looks like. There is no free trial. Plans are on the pricing page.

How to improve your visibility in Gemini

Tracking only helps if you act on it. These steps map to the two routes described above.

Fix what the index can see. Grounding uses Google Search. Make sure your key pages are indexed, load fast, and state plainly what you do, who it is for and how it compares. Do not block Googlebot. Note that the Google-Extended token controls use of your content for Gemini training, and does not affect Search crawling. Decide that on purpose, and check your robots.txt.

Write pages that answer the prompts. Look at the prompts where you are missing. Build a page for each real question: a comparison, a use case, a pricing explainer. Put the direct answer first, then the detail. Gemini tends to pull from pages that state facts in a clean way.

Get covered by the sources it already cites. Read the citation lists from your tracking. If the same review sites, directories or forums appear again and again, those are the places to be listed and described accurately. Your own site is rarely the only source.

Be consistent about your entity. Use the same name, description and category on your site, profiles and listings. Add structured data for your organisation and products. This helps Google connect the facts to one entity.

Correct errors at the source. If Gemini repeats an outdated price or feature, find where it came from. Usually it is an old page or listing. Update it, then watch the next runs.

Be patient with memory. Training data changes slowly. Retrieval changes fast. Expect grounded answers to respond to your work within weeks, and remembered facts to lag behind model updates.

A simple weekly routine

Run the fixed prompt set, three runs each.
Update mention rate, position, sentiment, citations and share of voice.
List the three biggest gaps: prompts where a competitor appears and you do not.
Ship one page or one listing fix aimed at each gap.
Review after four weeks.

Keep the changes small and log them, so you can tie a movement to an action.

Summary

Gemini answers from memory and from Google Search grounding. Track it with a fixed prompt set, repeated runs and five measures: mentions, position, sentiment, citations and share of voice. Improve it by getting your pages indexed and clear, by earning coverage on the sources it cites, and by fixing wrong facts where they start. Start by hand. Move to a tool when the volume gets in the way.

Originally published on pineprompt.com.

How to track your brand in Claude

Pineprompt — Fri, 16 Oct 2026 12:05:30 +0000

How to track your brand in Claude

Claude is Anthropic's AI assistant. People ask it for tools, vendors and comparisons, and it names brands. To track yours, pick a fixed set of prompts, run them on a schedule, and record whether Claude mentions you, where, how it describes you and which sources it cites.

This guide covers how Claude builds an answer, how to check it by hand, which prompts to track, what to measure and what you can do to improve the result.

How Claude sources its answers

Claude is a large language model made by Anthropic. It answers in one of two modes, and they give different results.

Web search off. Claude answers from its training data. It has a knowledge cutoff, so newer products, prices and news may be missing or out of date. There are no citations in this mode. If your brand is well covered in public text from before the cutoff, Claude may know it. If not, it may say it does not know you, or it may guess.

Web search on. Claude can search the web, read results and write an answer based on them. When it does, it shows the sources it used as citations, so you can see which pages shaped the answer. This is a form of retrieval-augmented generation. Claude decides for itself whether a question needs a search, and whether web search is available depends on the app, plan and settings. API use has its own tool setup.

So check both modes. Memory shows what Claude believes about you. Search shows whether your pages and the sites that cover you can be found today. See training data vs live retrieval.

Choose the prompts to track

Start with 20 to 50 prompts. Write them the way a buyer would ask, not the way a marketer would.

Category prompts: "What are the best project management tools for small agencies?"
Comparison prompts: "Compare Tool A and Tool B for a team of ten."
Problem prompts: "How do I reduce churn in a subscription app?"
Branded prompts: "What is Pineprompt and who is it for?"

Keep branded and non-branded prompts separate. Branded prompts show what Claude believes about you. Non-branded prompts show whether new buyers find you.

Freeze the wording. If you change a prompt, you lose the trend. Add new prompts as new rows. To find more phrasings buyers use, the prompt fan-out generator expands one prompt into the related searches an AI engine may run.

Check Claude manually

Open Claude in a new chat for every prompt, so earlier turns do not carry over.
Decide on web search for the run, on or off, and keep it the same across the set. Run the key prompts in both modes.
Record the date, the model name shown, and the prompt.
Copy the full answer into a sheet.
Note whether your brand appears, its position in any list, and how it is described.
If Claude searched, record each cited URL and its domain.
Run each prompt three times. Answers vary.

Twenty prompts at three runs each is sixty answers, and that is one snapshot. Repeat weekly.

What to measure

Mentions. The share of runs in which your brand is named. This is your brand mention rate. If it is zero, nothing else matters yet.

Position. Where you appear in a list or recommendation. See position in answer.

Sentiment. Whether the wording is positive, neutral or negative, and whether it is accurate. A confident false statement is a brand hallucination and needs a fix.

Cited sources. Which URLs Claude cites when web search is on. With search off there are none, so record that as a separate case. Separate your own pages from third party pages, and note the difference between being mentioned and being cited.

Share of voice. Your mentions divided by all brand mentions across your prompt set, compared with named competitors. See share of voice in AI.

Report each as an average across repeated runs. Expect answer volatility: a small weekly move is usually noise, while a steady change over a month is a signal.

Where manual tracking breaks down

Sixty answers a week is already an hour or more of copying and reading. Add competitors, more prompts and other engines and it stops being practical. Sheets also lose detail such as the source list and the exact wording.

Buyers use several assistants, and they disagree with each other. Pineprompt tracks six live AI platforms daily on every plan. Grok and Perplexity are temporarily unavailable. It stores the answers, mentions, position, sentiment and citations for each run. The Claude visibility tracker page shows what that looks like. There is no free trial. Plans are on the pricing page.

How to improve your visibility in Claude

Write pages that answer the prompts. For each prompt where you are missing, build a page for the real question: a comparison, a use case, a pricing explainer. Put the direct answer first.

Get covered by the sources it already cites. If the same review sites, directories or forums appear again and again in your citation lists, be listed and described accurately there.

Be consistent about your entity. Use the same name, description and category on your site, profiles and listings, and add structured data.

Act where the engine reads. Claude has no built-in index of its own that you can submit to. Search results come from the web, so crawlable, clear pages matter. Check that you are not blocking Anthropic's crawlers in your robots.txt by accident, and decide on purpose what you allow.

Correct errors at the source. If Claude repeats an outdated price or feature, find where it came from, update it, then watch the next runs.

A simple weekly routine

Run the fixed prompt set, three runs each.
Update mentions, position, sentiment, cited sources and share of voice.
List the three biggest gaps: prompts where a competitor appears and you do not.
Ship one page or listing fix for each gap.
Review after four weeks.

Summary

Track Claude with a fixed prompt set, repeated runs and five measures: mentions, position, sentiment, cited sources and share of voice. Improve it by publishing clear pages, earning coverage on the sources it cites, and fixing wrong facts where they start. Start by hand, and move to a tool when the volume gets in the way.

Originally published on pineprompt.com.

How to track your brand in ChatGPT

Pineprompt — Fri, 16 Oct 2026 08:06:46 +0000

Someone types "what is the best project management tool for a small agency" into ChatGPT. It names three products. Yours is either one of them or it is not.

You cannot see that in Google Analytics. You cannot see it in Search Console. If you want to track brand mentions in ChatGPT, you have to ask ChatGPT, and you have to do it in a way that produces numbers you can trust.

This guide walks through how. Start by hand. Then decide when hand-work stops being worth it.

Why ChatGPT answers are hard to track

A search ranking is stable enough to check once a day. A ChatGPT answer is not. Three things get in the way.

Answers vary. Ask the same question twice and you may get a different list of brands, in a different order. One run tells you almost nothing. You need repeated runs to see how often you appear.

Prompts vary. Your buyers do not all phrase the question the same way. "Best CRM for startups" and "which CRM should a five-person startup use" can return different brands.

Context varies. Whether ChatGPT searches the web for an answer changes what it can cite. Answers built from search can link to sources. Answers built from the model alone may only name brands.

So the job is not "check if we show up." The job is "measure how often we show up across a fair set of prompts, over time."

Step 1: Build a prompt set

Write down the questions your buyers actually ask. Aim for 20 to 50 to start. Cover four types:

Category prompts. "Best [category] tools." These decide who gets shortlisted.
Use-case prompts. "[Category] for [audience or job]." These are narrower and often easier to win.
Comparison prompts. "[Your brand] vs [competitor]." These show how you are framed.
Problem prompts. "How do I [solve the problem you solve]?" These catch buyers before they know your category name.

Pull the wording from sales calls, support tickets, search queries and community threads. Real phrasing beats invented phrasing.

If you are short on ideas, the prompt fan-out generator takes one seed prompt and expands it into the related sub-questions AI platforms tend to run behind the scenes. It is a fast way to widen a prompt set beyond the obvious ten.

Step 2: Run each prompt more than once

Open a fresh chat for every run. Old conversation context bleeds into later answers, and you want each test clean.

Run each prompt at least five times. Ten is better for the prompts that matter most. Log every run. You are not looking for one answer. You are looking for a rate.

Two practical notes:

Test with and without web search when the option is available to you. They are different behaviours and may produce different brands.
Log the date. Answers change as models and sources change, so a result from March says little about September.

Step 3: Record the right fields

For every run, capture:

Field	What it tells you
Prompt	Which question was asked
Brand mentioned (yes/no)	Whether you are in the answer at all
Position	First, second, third, or later in the list
Sentiment	How the answer describes you
Competitors named	Who you are sharing the answer with
Sources cited	Which pages the answer linked to

A spreadsheet is enough at this stage. One row per run.

The sources column is the one people skip and the one that pays off. If ChatGPT keeps citing the same review site or the same competitor comparison page, that page is shaping your category. It is where your work should go.

Step 4: Turn rows into metrics

Raw rows are noise. Turn them into three numbers.

Visibility rate. Runs where your brand appeared, divided by total runs. If you appeared in 12 of 40 runs, your visibility rate is 30%.

Share of voice. Your mentions divided by all brand mentions across the same runs. This tells you how you compare with competitors, not just how you do alone.

Average position. Where you land when you do appear. Being named first is worth more than being named fifth.

Track each by prompt group, not only overall. A strong overall number can hide a category prompt where you never appear.

Step 5: Read the sources, not just the mentions

A mention is the outcome. A citation is the cause.

When ChatGPT answers with web search, it often links the pages it used. Collect those links. Sort them into types: your own site, competitor sites, review platforms, forums, news, listicles.

Then ask two questions.

Which sources show up for prompts where we win? Protect and extend those.
Which sources show up where we lose? These are your targets. Get listed, get reviewed, or publish something better.

This is where tracking turns into a plan. You are no longer guessing why you are missing. You are looking at the pages the answer was built from.

Step 6: Set a cadence

Run the full prompt set on a fixed schedule. Weekly works for most teams. Monthly is the floor. Anything less and you will miss a change until it has already cost you.

Keep the prompt set stable between runs. If you change the prompts every time, you cannot tell whether the answers moved or you did. Add new prompts, but keep a core set fixed for trend lines.

Where manual tracking breaks down

Manual tracking works. It also gets expensive fast.

Do the arithmetic. 40 prompts, 5 runs each, is 200 chats per pass. Weekly, that is 800 a month, for one platform. Add a second platform and you have doubled it. Add logging, sorting sources and building charts, and it becomes somebody's job.

It also stays stuck in ChatGPT. Your buyers use several assistants. Gemini, Perplexity, Claude, Google AI Overviews, AI Mode, Grok and Copilot each pull from their own mix of sources, and a brand that wins on one can be invisible on another. Tracking one engine gives you a partial picture.

A rule of thumb: do it by hand until you have a prompt set you believe in and you know which metrics you care about. Then automate it.

What to do with the results

Tracking is only useful if it changes what you do. A few common moves:

Fix missing basics. If you never appear for a use case, check that you have a page that answers it plainly.
Get into the cited sources. If a comparison article dominates citations, work out how to be included in it.
Correct wrong descriptions. If ChatGPT gets your pricing or features wrong, find the page it learned that from and fix or outrank it.
Watch competitors. Sudden jumps in a rival's visibility usually trace back to a specific piece of content or coverage.

Retest after each change. That is the loop: measure, change one thing, measure again.

Track it automatically

If you would rather not run 800 chats a month yourself, Pineprompt does it for you. It runs your prompts on a schedule across ChatGPT and seven other AI engines, then reports visibility, share of voice, position, sentiment and the sources behind each answer. There is a page on ChatGPT visibility tracking that shows what that looks like. Plans start at $99 a month with every engine included, and the details are on the pricing page.

Either way, start this week. Write 20 prompts, run each five times, and see where you stand. You will learn more from that one spreadsheet than from any amount of guessing.

Originally published on pineprompt.com. Read the original.

What is prompt fan-out?

Pineprompt — Fri, 16 Oct 2026 08:06:10 +0000

Prompt fan-out is what happens when an AI search engine takes one question and turns it into several smaller searches. You ask one thing. The system searches for many related things, then combines what it finds into one answer.

For anyone trying to be visible in AI answers, this matters. Your page may never match the question the buyer typed. It only has to match one of the sub-queries. The glossary entry on prompt fan-out has the short definition. This guide goes further.

How fan-out works

Take a buyer who types: "best accounting software for a small design studio."

A person would search that phrase and scan the results. An AI system often does something different. It may break the question into parts such as:

accounting software for freelancers
accounting software with invoicing
pricing of popular accounting tools
reviews of accounting software from small businesses
accounting software integrations with payment tools

It runs those searches, reads what comes back, and writes one answer. The sources cited in that answer may come from any of the sub-queries, not from the original question.

The exact sub-queries are chosen by the system, and they can change from run to run. You usually cannot see all of them.

Who does it

Google has described this technique for AI Mode. Its documentation says AI Mode uses a "query fan-out" approach, issuing multiple related searches across subtopics and data sources. Google says AI Overviews can use the same technique.

Other assistants also search the web when a question needs fresh information. Microsoft's documentation for Copilot, for example, says it identifies terms where a search would improve the response and generates a search query. Whether a given assistant issues one query or many differs by product and by question, and it changes over time. Treat fan-out as a common pattern, not a fixed rule for every engine.

Why it matters for visibility

You compete on more than one query

A single question can lead to five or ten searches. Each is a chance to be found or missed. A page that ranks for a narrow sub-query can be pulled into an answer for a broad question.

Broad pages can lose to specific ones

A general page about "accounting software" may not match "invoicing features for freelancers" as well as a page written on exactly that. Specific, well-structured pages give the system something precise to pick up.

Your brand can be missing from part of the picture

If your product is strong on invoicing but no page or third-party source says so, the invoicing sub-query will not find you. You may be absent from the answer without knowing why.

Cited sources become clearer

The sources in an answer often reflect the sub-queries. Reading them tells you which angles the system explored, and which sites it trusted for each.

How to use fan-out to pick prompts to track

You cannot track every possible question. Fan-out gives you a way to choose a set that covers real ground.

Step 1: Start with the buyer questions

List the questions buyers ask before they pick a product like yours. Include category questions, comparisons and problem questions. Aim for 10 to 15 to begin with.

Step 2: Expand each one into sub-queries

For each question, ask what a system would need to know to answer it. Look for:

Attributes: price, features, size of company it suits, integrations.
Comparisons: you against named alternatives.
Proof: reviews, case studies, independent write-ups.
Use cases: the jobs people hire the product for.
Constraints: country, language, budget, compliance.

You can do this by hand. Or use the free Prompt fan-out generator: enter a question and it suggests related sub-queries you can test.

Step 3: Turn the good ones into prompts

Not every sub-query makes a good tracking prompt. Keep the ones a real buyer might ask an assistant in natural language. Drop duplicates and anything too vague to give a stable answer.

Step 4: Group them

Keep prompts in groups, for example category, comparison, use case and pricing. When results come in, you can see which group is weak instead of staring at one blended number.

Step 5: Fix the list and track it

Keep the set stable so you can compare over time. Add new prompts as a separate group. Each prompt should be run repeatedly, since answers vary.

What to measure once you are tracking

For each prompt, record:

Mention: is your brand named?
Position: where in the answer does it appear?
Sentiment: how are you described?
Cited sources: which of your pages and which other domains are linked?
Competitors: who else is named, and how often?

Share of voice ties these together: your mentions as a share of all brand mentions across the set.

Read the results by group. If you are named on category prompts but not on pricing prompts, the gap is in pricing content or in what third parties say about your pricing. That is a concrete task, which a blended score would hide.

A worked example

Suppose you sell scheduling software for clinics. Your seed question is "best appointment scheduling software for a small clinic."

Expanded, it might give these prompts:

Group	Prompt
Category	best appointment scheduling software for small clinics
Attribute	scheduling software with automatic patient reminders
Comparison	your product vs a named competitor for clinics
Constraint	scheduling software that meets health data rules in Germany
Use case	how to reduce missed appointments at a clinic

Five prompts from one question. Each can trigger its own sub-queries. Track all five and you will see whether you show up across the range, or only on the obvious one.

What to do with the findings

Missing on an attribute: publish a clear page that covers it, in plain language.
Missing on comparisons: write an honest comparison page. Check what third-party comparisons say about you.
Cited sources are all third parties: work on being present and accurately described there.
Wrong or dated descriptions: fix your own pages first, then the sources the answers cite.

None of this is a trick. It is ordinary content work, aimed by better information.

Limits to keep in mind

You cannot see every sub-query. Generators and manual expansion give you likely ones, not the exact list a system ran.

It changes. Systems are updated. A pattern that held last quarter may not hold now.

It is not a guarantee. Covering a sub-query well raises your chance of being used. It does not force a mention.

Tracking across engines

Fan-out affects AI Mode, AI Overviews and other assistants that search the web, so track across them. See our pages on Google AI Mode tracking and Google AI Overview tracking for the Google side.

Pineprompt tracks prompts daily across ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, Google AI Mode, Grok and Copilot, with country and language targeting. It reports mentions, position, sentiment, cited sources, competitors and share of voice. The methodology page explains how runs are recorded. Plans start at $99 a month with no free trial; see pricing.

The short version

One question becomes many searches. Your visibility depends on the sub-queries, not only the headline question. Use fan-out to widen your prompt list, group the prompts, keep the list fixed, and read results by group. Start with one buyer question and the generator today.

Originally published on pineprompt.com. Read the original.

How to get your brand cited by Perplexity

Pineprompt — Fri, 16 Oct 2026 07:08:22 +0000

Most AI assistants give you an answer and hide the work. Perplexity shows it. Every answer comes with numbered sources, and you can click each one.

That is good news for brands. You can see exactly which pages Perplexity used, which means you can see what it takes to be one of them.

This guide covers what to fix, what to write and how to check that it worked.

How Perplexity builds an answer

Perplexity is a search-first product. When you ask a question, it searches the web, reads the results and writes a response that links back to the pages it relied on.

Two consequences follow.

You need to be findable. If Perplexity cannot crawl your pages, it cannot cite them. That is a technical question and the cheapest one to fix.

You need to be worth citing. Among the pages it can reach, it prefers ones that answer the question clearly and come from sources that look credible. That is a content question.

Nobody outside Perplexity knows the exact ranking rules, and anyone who claims to is guessing. What you can do is remove obstacles, publish pages that suit how answers are assembled, and measure results.

Step 1: Let PerplexityBot in

Perplexity documents two user agents:

PerplexityBot surfaces and links websites in Perplexity search results. Perplexity says it is not used to crawl content for AI foundation models.
Perplexity-User supports user actions. When a person asks a question, Perplexity may visit a page to answer it. Perplexity says this fetcher generally ignores robots.txt rules, because a user requested the fetch.

Perplexity recommends allowing PerplexityBot in robots.txt and permitting requests from its published IP ranges. The second part matters. Many sites block bots at the firewall or CDN without knowing it, and a clean robots.txt does not help if a WAF returns a 403.

Check both. Our crawler access guide covers the robots.txt side in detail. Perplexity also says changes can take up to 24 hours to take effect.

Step 2: Check the basics of your pages

Before writing anything new, run your key pages through the GEO readiness checker. It looks at the practical foundations that make a page easy for an AI engine to read and use, and shows you what is missing.

The usual gaps are dull and fixable:

Content that only appears after JavaScript runs
Pages with no clear title, headings or summary
Thin pages that say little
Missing or broken structured data
No obvious signal of who wrote the page or who runs the site

Fix these first. They cost little, and they apply to every engine, not just Perplexity.

Step 3: Write pages that answer the question

Perplexity pulls passages, not whole pages. A passage that answers a question in a few plain sentences is easier to use than one buried in a long story.

Some habits that help:

Lead with the answer. Put the direct answer in the first lines under a heading, then explain. Do not make the reader wade through a preamble.

Use question-shaped headings. "How much does X cost?" beats "Pricing philosophy." Headings that mirror real questions match how people prompt.

Be specific. Name the product, the numbers, the limits, who it is for and who it is not for. Vague copy gives an engine nothing to quote.

Keep facts current and consistent. If your pricing page says one thing and a blog post says another, the engine may pick either. Or neither.

Cover the comparison. Buyers ask "X vs Y" constantly. A fair, factual comparison page on your own site gives you a chance to control that framing.

Step 4: Earn mentions elsewhere

For many category questions, Perplexity cites third-party pages, not vendor sites. Think review platforms, comparison articles, forum threads, industry publications and directories.

If those pages do not mention you, Perplexity has little reason to. So look at who does get cited and work backwards:

Run your category prompts in Perplexity.
List every source that appears in the citations.
Sort them into own site, competitor sites, reviews, forums, news and lists.
Pick the third-party pages that appear most often.
Work out how to be included: a listing, a review, a briefing, a contribution, a correction.

This is slower than editing your own site. It is also where much of the ground is won. Do it honestly. Ask for accurate coverage, not favourable fakes.

Step 5: Keep the site trustworthy

Engines lean on credibility signals, though the specifics are not public. The sensible things are the same ones that help any reader:

Named authors with real credentials
An about page that says who you are
Accurate, dated content that you actually maintain
Original data, examples or tooling that others reference

Do not invent any of it. Fabricated authors and made-up numbers are easy to contradict, and they harm you with readers well before any engine notices.

Step 6: Measure citations over time

Perplexity makes measurement easy because the sources are on screen. Build a prompt set of 20 to 50 real buyer questions. Run each one several times, since answers vary between runs. Record:

Whether your brand is named
Whether your domain is cited, and which page
Where you appear in the source list
Which competitors and third-party sites appear alongside you

Then compute a citation rate: runs where your domain was cited, divided by total runs. Track it by prompt group and compare month to month.

If you need prompt ideas, the prompt fan-out generator expands a seed question into the related sub-queries that tend to sit behind an answer.

Change one thing at a time. If you rewrite a page and add three listings in the same week, you will not know what moved the number.

Mentions and citations are different

A citation means your page was linked as a source. A mention means your brand was named in the text. You can have one without the other. Perplexity might cite your blog post while recommending a competitor, or name you without linking you.

Track both. A citation with no mention is traffic potential. A mention with no citation is awareness you do not control.

Do not chase tricks

There are plenty of promised shortcuts: hidden text, keyword stuffing, bulk-generated pages. Skip them. Perplexity is built on search, and search has spent decades learning to discount that kind of thing. Pages that are useful to a human reader are the ones worth citing.

Make it routine

Perplexity citation work is not a project. It is a loop: check access, improve pages, earn mentions, measure, repeat. Monthly is a reasonable rhythm for content changes, and weekly for measurement.

Pineprompt automates the measurement part. It runs your prompts on Perplexity and seven other engines, and reports mentions, citations, position and sources. See how it works on the Perplexity visibility tracker page. Plans start at $99 a month with all engines included, listed on the pricing page.

Originally published on pineprompt.com.

Should AI crawlers access your site?

Pineprompt — Fri, 16 Oct 2026 07:07:53 +0000

Many sites added Disallow: / for GPTBot in 2023 and never looked at it again. Some went further and blocked every AI user agent they could find.

That is a legitimate choice. But it is often made without knowing that the crawlers do different jobs. Block the training bot and you keep your content out of training. Block the search bot and you may vanish from AI answers.

This guide separates the two so you can decide on purpose.

The core idea: one company, several bots

OpenAI, Anthropic and Perplexity each run more than one user agent. Their own documentation splits them by purpose, and each is controlled separately in robots.txt.

OpenAI

Per OpenAI's crawler documentation, there are these robots:

GPTBot crawls content that may be used to train OpenAI's generative AI foundation models. Disallowing it signals that your content should not be used for training.
OAI-SearchBot surfaces websites in ChatGPT's search features. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links.
ChatGPT-User visits pages for certain user actions in ChatGPT and Custom GPTs. OpenAI says it is not used to crawl automatically, and because the actions are user-initiated, robots.txt rules may not apply.
OAI-AdsBot validates pages submitted as ads on ChatGPT. It is not used to train models.

OpenAI states that each setting is independent. You can allow OAI-SearchBot and disallow GPTBot. It also says robots.txt changes can take about 24 hours to take effect for search.

Anthropic

Anthropic's help center lists three:

ClaudeBot collects web content that could contribute to model training. Restricting it signals that your future content should be excluded from training datasets.
Claude-User fetches pages when a person asks Claude a question. Disabling it may reduce your visibility for user-directed web search.
Claude-SearchBot navigates the web to improve search result quality. Disabling it prevents indexing for search optimization and may reduce your visibility in user search results.

Anthropic says its bots honor robots.txt and support the non-standard Crawl-delay directive. It also warns that blocking by IP address may not work reliably, because it stops Anthropic reading your robots.txt.

Perplexity

Perplexity's documentation lists two:

PerplexityBot surfaces and links sites in Perplexity search results. Perplexity says it is not used to crawl content for AI foundation models.
Perplexity-User supports user actions. Perplexity says it generally ignores robots.txt rules, since a user requested the fetch.

Training bots and search bots are a different decision

Here is the useful way to think about it.

Training crawlers (GPTBot, ClaudeBot) feed future models. Blocking them is about control over how your content is used. It has no documented effect on whether you appear in live search answers.

Search crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) feed the retrieval side. Blocking them takes you out of the pool of sources those products can surface and link.

User-triggered fetchers (ChatGPT-User, Claude-User, Perplexity-User) act on behalf of one person asking one question. Some vendors say robots.txt may not apply to them.

If your goal is to be recommended by AI assistants, allow the search crawlers. Whether to allow the training crawlers is a separate business call.

How to decide

Ask three questions.

1. Do we want AI assistants to recommend us? For most businesses selling something, yes. Allow the search bots.

2. Is our content the product? Publishers, paid research firms and licensed-content sites may reasonably block training bots. Their content is the thing they sell.

3. Are we comfortable with training use? For a software company with public marketing pages, many teams decide the upside of being known to models outweighs the cost. Others disagree. Both are defensible. Note that blocking training bots only affects future crawling, not data already collected.

There is no universal right answer. There is only an answer you chose knowingly.

Example robots.txt files

Each group starts with a User-agent line. Rules apply to the agent named in that group.

Allow everything AI

User-agent: GPTBot
Allow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: ChatGPT-User
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: Claude-User
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Perplexity-User
Allow: /

If you have no rule blocking these agents, they are generally allowed by default. Listing them is optional but makes intent explicit.

Appear in AI search, opt out of training

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

This matches each vendor's own description: training bots blocked, search bots allowed. Be aware that it is your reading of their documentation. Vendors can change their bots, so recheck now and then.

Block a private area only

User-agent: GPTBot
User-agent: ClaudeBot
User-agent: PerplexityBot
Disallow: /account/
Disallow: /internal/

Stacking several User-agent lines above one set of rules is valid. Most sites should do this rather than block everything.

Common mistakes

A blanket User-agent: * block. This blocks every bot without a more specific group, including all the search crawlers above. Check any wildcard rules.

Blocking at the firewall. A CDN or WAF bot rule can return 403 to crawlers that robots.txt allows. Perplexity and OpenAI both publish IP ranges so you can allowlist them. Anthropic asks that you do not rely on IP blocking to opt out.

Typos in user-agent names. The names are GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, Claude-User, PerplexityBot and Perplexity-User. A misspelled name matches nothing and fails silently.

Assuming robots.txt stops everything. It is a request, not a lock. Vendors say their bots honor it, and some user-triggered fetchers may not. If content must be private, put it behind a login.

Forgetting subdomains. Robots.txt applies per host. Anthropic notes you should do this for every subdomain.

Test what you set up

Do not trust a file you have not tested. The AI crawler access checker reads your robots.txt and shows which AI user agents are allowed or blocked. Run it after every change, and again after any CDN or firewall update.

Then check your server logs for the user-agent names above. If you allowed them and never see them, something between the internet and your server is in the way.

Also consider llms.txt

Robots.txt says what bots may fetch. llms.txt is a proposed plain-text file that points language models to your most useful pages. Treat it as a low-cost extra, not a guaranteed fix. If you want one, the llms.txt generator builds a starting file from your site.

Then measure the outcome

Changing robots.txt is an input. What you care about is the output: do AI assistants name and cite you? After you settle your rules, track it. Pineprompt runs your prompts across ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, AI Mode, Grok and Copilot, and reports where you appear and which sources are cited. Plans start at $99 a month with every engine included. See the pricing page.

Originally published on pineprompt.com.

Pineprompt AI search visibility benchmark 2026

Pineprompt — Thu, 15 Oct 2026 14:25:23 +0000

Pineprompt analyzed 18,792 responses to 200 non-branded buyer prompts across 10 categories, run daily on seven AI search engines from September 30 to October 13, 2026. This is how the engines cite sources and name brands.

Key findings

Perplexity cites a source in 100% of its answers (Gemini: 90.7%).
The average answer names 6.2 tracked brands.
12.9% of brands are named by every engine; 46.2% by only one.
Perplexity averages 10.1 citations per answer.
Perplexity shows the highest day-to-day brand churn, at 68.8%.

The data is licensed under CC BY 4.0, so you may share and adapt it with attribution. Source: Pineprompt AI Search Visibility Benchmark 2026, pineprompt.com/research/ai-search-visibility-benchmark. The CSV is available.

Results by engine

Engine	Answers with citations	Avg citations per answer	Avg brands named per answer	Day-to-day brand change
ChatGPT	99.5%	9.3	5.5	36%
Claude (claude-haiku-4-5-20251001)	95.3%	4.4	6.7	46.4%
Claude (claude-haiku-5-5)	96.4%	5	8.5	41.1%
Copilot	98%	4.8	7.6	51.3%
Gemini	90.7%	6.7	6	55.8%
Google AI Mode	94.8%	5.9	6.5	45.6%
Google AI Overview	96.3%	7.1	5.7	45.1%
Perplexity	100%	10.1	3.9	68.8%

Top cited domain types by engine

Engine	Most cited domain types, share of citations
ChatGPT	Vendor and brand sites (48.8%); Editorial and review sites (44.5%); Other (5.6%)
Claude (claude-haiku-4-5-20251001)	Editorial and review sites (51.9%); Vendor and brand sites (47.6%); Other (0.3%)
Claude (claude-haiku-5-5)	Editorial and review sites (52.7%); Vendor and brand sites (46.9%); Other (0.3%)
Copilot	Editorial and review sites (56.6%); Vendor and brand sites (41.9%); Other (0.7%)
Gemini	Vendor and brand sites (59.8%); Editorial and review sites (38.4%); Forums and communities (1.2%)
Google AI Mode	Vendor and brand sites (34.5%); Editorial and review sites (27.5%); Video platforms (18.5%)
Google AI Overview	Vendor and brand sites (39.4%); Editorial and review sites (28.5%); Video platforms (18.5%)
Perplexity	Vendor and brand sites (46.3%); Editorial and review sites (43.9%); Other (5.4%)

Brand overlap between engines

12.9% of the 8,269 brands seen are named by every engine.
46.2% are named by only one engine.
Google AI Overview and ChatGPT share 45.6% of named brands (Jaccard similarity).
Google AI Mode and ChatGPT share 45.2% of named brands (Jaccard similarity).

Results by category

Category	Answers with citations	Avg citations per answer	Avg brands named per answer	Day-to-day brand change
Accounting	97.1%	6.9	6.2	41.4%
CRM	96.5%	6.6	5.9	48.9%
Cybersecurity and VPN	97.1%	6.9	5.4	43.1%
Ecommerce platforms	97.9%	7.3	6.7	44.1%
Email marketing	97.3%	6.8	6.1	40.9%
HR and recruiting	98%	6.5	6.7	47%
Project management	95.7%	6.6	6.6	44.1%
Running and fitness	91.2%	6.5	5.1	46.8%
Travel insurance and travel	94.1%	6.8	5.9	43.9%
Web hosting	97.7%	6.8	6.9	47.4%

Methodology

We wrote 200 non-branded buyer prompts across 10 categories and ran each against 7 engines: ChatGPT, Gemini, Perplexity, Claude, Google AI Overview, Google AI Mode, and Copilot. The 18,792 responses were collected from 2026-09-30 to 2026-10-13. Prompts were scheduled daily on each engine for 14 days (US, English); some runs returned no answer. Pineprompt analysed the responses to extract citations and named brands. Metric definitions are on the methodology page.

Limitations

The prompts are a fixed sample of buyer questions in US English, so results may not hold for other markets, languages, or query styles. AI answers are non-deterministic, and engines change without notice, so the numbers describe a 14-day window rather than a lasting property of each engine. Brand stats cover brands seen with 2 or more mentions that were assigned to category segments by 2026-10-13. Claude is split by model: the previous Haiku until 2026-10-09 around 18:40 UTC, claude-haiku-5-5 after. Domain type uses the app's source classification; missing source types land in Other.

Originally published on pineprompt.com.

How to measure share of voice in AI search

Pineprompt — Thu, 15 Oct 2026 14:23:43 +0000

Share of voice in AI search is your brand's portion of all brand mentions in AI answers to a fixed set of prompts. If assistants name four brands a total of 400 times and yours 60 times, your share is 15%. The number only means something if the prompts, competitors and method stay fixed.

This guide covers the definition, how to build the prompt and competitor sets, three ways to count, how many runs you need, why per-engine and blended figures differ, how country and language change results, a worked example and common mistakes.

The definition

Share of voice = your brand mentions divided by all brand mentions in a fixed prompt set, across the brands you chose to compare.

Two things follow. The result depends on who is in the denominator, so the competitor set is part of the definition. And it is relative: your share can fall while your mentions rise, if competitors grow faster. For the short version, see share of voice in AI. Pineprompt's share of voice scoring feature and the metric docs describe how it is calculated there.

Choose the prompts

Use 30 to 100 prompts that a buyer would type. Mix three kinds:

Category prompts: "What are the best invoicing tools for freelancers?"
Comparison prompts: "Compare Tool A and Tool B for a small team."
Problem prompts: "How do I stop chasing late invoices?"

Keep branded and non-branded prompts apart. A prompt that names your brand will inflate your share. Measure share of voice on non-branded prompts, and use branded ones for a separate check on what assistants say about you.

Freeze the wording. If you change a prompt, start a new series. The prompt fan-out generator can help you find more phrasings for the same need.

Choose the competitor set

List 3 to 8 brands that buyers actually consider, not only the ones you consider rivals. Take them from the answers themselves: run the prompts first, and note which brands appear. Add the ones that show up often.

Decide up front how to treat brands outside your list. Either ignore them, or group them as "other". Pick one rule and keep it. Adding or removing a competitor changes everyone's share, so record the date of any change.

Three ways to count

Mention share. Every named brand counts once per answer. This is the simplest and the one used in the definition above. It treats the first recommendation and the fifth as equal.

Position-weighted share. Give more weight to brands listed earlier, for example 3 points for first, 2 for second and 1 for any later place. Share is your points divided by all points. It rewards being recommended first. See position in answer.

Citation share. When an engine shows sources, count how often your domain is cited compared with competitor domains. This measures who supplies the evidence, which is not the same as who is named. See mention vs citation.

Report all three if you can. They answer different questions: are we named, are we named early, and are our pages used.

Sampling and repeated runs

AI answers change from run to run, even for the same prompt on the same day. This is answer volatility. One run per prompt gives a noisy number.

Run each prompt at least three times per period, and more for the prompts that matter most.
Start each run in a new chat so earlier turns do not carry over.
Average across runs. Count the share per run, then average, or pool all mentions. Pick one method.
Treat small weekly moves as noise. A steady change over four weeks or more is a signal.

If you track daily, compare rolling averages, not single days.

Per-engine and blended figures

Engines name different brands. A brand can have a strong share in one assistant and almost none in another. A single blended figure hides that.

Report share for each engine first. Then, if you need one number, blend them with a stated rule: an equal average, or weights based on where your buyers actually search. Whichever you choose, keep it fixed and show the per-engine figures beside it. Engines that cite sources and engines that answer from memory also behave differently, so do not mix them up when reading citation share.

Country and language

Answers depend on location and language. A prompt in English asked from the US can name different brands than the same prompt in German asked from Germany. Local brands appear in their home market, and some engines show different answers or sources by region.

Treat each country and language pair as its own series, with its own prompts written natively, not machine translated. Do not average markets together unless that is the question you are asking.

A worked example

All names and numbers here are hypothetical. Larkspur is a fictional invoicing tool compared with three fictional rivals. The set is 50 non-branded prompts, run 3 times on each of two engines, A and B. That gives 150 answers per engine.

Mention counts across all answers:

Brand	Mentions	Mention share
Larkspur	60	15%
Orbitly	140	35%
Quillnet	120	30%
Tandem Works	80	20%
Total	400	100%

Now weight by position (3, 2, 1 points). Larkspur is usually listed late, so its points are 110 out of 810, a share of 13.6%. Orbitly, often listed first, has 330 points, or 40.7%. Quillnet has 240 (29.6%) and Tandem Works 130 (16.0%). Larkspur looks slightly worse once position counts.

Citation share, from 200 cited links to these four brands' domains: Larkspur 20 (10%), Orbitly 70 (35%), Quillnet 70 (35%), Tandem Works 40 (20%).

Per engine, Larkspur has 50 of 200 mentions on engine A, which is 25%, and 10 of 200 on engine B, which is 5%. The blended equal average is 15%, which matches the pooled figure but hides a ten-fold gap between engines.

What this tells Larkspur: it is named in some places but late, rarely cited, and nearly absent on engine B. That is three different problems.

Common mistakes

Changing the prompt or competitor set mid-series. The trend then reflects your edits.
Including branded prompts. They push your share up without telling you about new buyers.
One run per prompt. You will chase noise.
Only a blended number. It hides the engine where you are missing.
Mixing markets. Different countries and languages are different series.
Treating mention share as sales. It measures visibility in answers, not revenue.
Counting a brand name inside a negative or wrong sentence as a win. Check how you are described; a wrong claim is a brand hallucination.

How to act on the results

Start with prompts where competitors appear and you do not. Each is a content or coverage gap.
Look at the engines with the lowest share, then at the sources they cite. If the same review sites or directories keep appearing, get listed and described accurately there.
If your mention share is fine but position is low, improve how clearly your pages state who you are best for.
If mentions are high but citations low, publish pages that answer the prompt directly, with facts that are easy to quote.
Make one change at a time where you can, and compare after at least four weeks.

Doing this at scale

By hand, 50 prompts at three runs on several engines is hundreds of answers per period. Pineprompt tracks six live AI platforms daily. Grok and Perplexity are temporarily unavailable. It stores each answer and calculates share of voice for you. The methodology page explains how runs are made and counted. There is no free trial; see pricing for plans.

Summary

Share of voice is your mentions divided by all brand mentions in a fixed prompt set. Fix the prompts and competitors, repeat runs, report per engine and per market, and look at mention, position and citation share together. Then act on the specific gaps the numbers show.

Originally published on pineprompt.com.