Most "GEO Best Practices" Are Wrong

zheng zhongwei — Fri, 04 Sep 2026 08:09:52 +0000

I spent the last two weeks doing something I'd been putting off: reading every large-sample AI search study published in the past six months. Not the "AI is coming, here's what to do" marketing fluff — I mean studies with clear sample sizes, statistical methods, and reproducible results.

I started with ~100 pieces. Only 26 made the cut, spanning December 2025 to August 2026. The rest were either tool tutorials or thinly-veiled product pitches.

This post distills the most valuable conclusions from those 26 reports, each tagged with its sample size. Many contradict the "GEO best practices" you've been sold — which is exactly why they're worth sharing.

Disclaimer: personal synthesis + cross-validation. Not infallible. Happy to debate.

What This Post Covers

Six hard findings — if you only read one section, read this one
Debunked tactics — these "GEO service packages" need to stop
Page-level writing framework — no tools needed, just change how you write
Technical accessibility — if AI can't crawl you, nothing else matters
Fan-out queries — AI search's real unit of topic planning
Off-site signals — AI learns about you from everyone else
Measurement — track average visibility, not single-shot answers
90-day execution checklist

1. Six Hard Findings

These six come from six independent large-sample studies. They cross-validate each other. If you only read one screen, read this one.

YouTube Mentions Are the Strongest Single Factor (correlation 0.740)

A study of 75,000 brands measured Spearman correlation between various factors and AI visibility. YouTube mentions scored 0.740 in AI Mode, 0.737 in ChatGPT, and 0.712 in AI Overviews — outranking everything else.

And mention count matters more than view count — a mention in a low-view video counts just the same.

Page Count Barely Moves the Needle (correlation 0.194)

Same 75K-brand study: the correlation between total indexed pages and AI visibility is just ~0.194. Domain Rating only reaches 0.266–0.326. The old playbook of "publish more + build DR" doesn't convert in AI search.

Get Indexed First, Then Get Cited (88.46%)

A study of 1.4 million ChatGPT prompts found that URLs retrieved via the search channel had an 88.46% citation rate. News dropped to 12.01%, Reddit to just 1.93%. Traditional SEO isn't the opposite of AEO — it's the prerequisite.

Citations Cluster in the Top Third (44.2%)

Across 1.2M ChatGPT citations, 44.2% referenced content in the first 30% of the page, 31.1% the middle, 24.7% the end. Bury your answer in paragraph 12 and your citation probability drops to roughly 1/2.5.

AI Overviews No Longer Favor the Top 10 (only 37.9%)

Across 863K SERPs and 4M AIO citations, only 37.9% of cited URLs appeared in the top 10 organic results. Positions 11–100 accounted for 31.2%, and beyond 100 accounted for 31.0%. Six months ago, the top-10 share was ~76%.

AI Traffic Is Tiny but Converts at 23x

One platform shared its own data: AI search visitors made up just 0.5% of traffic but drove 12.1% of signups — a 23x conversion rate over organic search. An independent SaaS reported 20.15% vs 7.06%. Stop judging this channel by traffic share.

2. Debunked Tactics

This section is worth more than the positive advice: every item below is actively being sold as a "GEO service package" — and every one has been disproven by large-sample data. Stop these first.

✕ "Adding Schema Markup Boosts AI Citations"

The correlation looks compelling: 53% of cited pages had schema; cited pages were ~3x more likely to have JSON-LD than uncited pages. But matched difference-in-differences killed the causal claim: tracking 1,885 pages that added JSON-LD against 4,000 controls yielded AI Overviews −4.6% (statistically significant), AI Mode +2.4%, ChatGPT +2.2% (indistinguishable from zero). A separate packet-capture study confirmed none of the five major AI systems actually parse schema during live retrieval.

→ Keep schema for rich results and Knowledge Graph. Don't budget it as an AEO lever.

✕ "Deploy llms.txt to Help AI Read Your Site"

A scan of 137,210 domains: 28% had published a valid llms.txt, but 97% received zero requests during the study period. The remaining 3% captured all ~22,000 requests — 96% from bots. Real AI retrieval bots accounted for just 1.1%; the largest traffic source was SEO audit tools (21.7%).

→ Only worth it for developer documentation sites targeting coding agents. For marketing sites, let your CMS auto-generate it.

✕ "Publish More Pages / Raise DR to Get Into AI Answers"

Page count correlation: 0.194. DR: 0.266–0.326. Cross-check: Wikipedia averages 6.69 citations per page vs. 0.48 for a content clone site. The clone had only 5x fewer pages but 70x fewer citations — with an average of 68 outbound reference links per page. Volume doesn't work.

→ Your north-star metric is citations-per-page, not page count.

✕ "Create 'Best Tools' Listicles and Rank Yourself #1"

A test across 5 domains, 34 self-promotional pages, tracking 9,886 AI responses over 4 months: effective for brand-new brands (captured 72 previously empty prompt×engine combinations, 82% citing the self-promotional page). But nearly useless for brands AI already knows — only 6% of new mentions came from self-promotional pages vs. 94% from third-party content. Worse: 43% of AI responses citing the self-promotional page didn't mention the target brand at all (control group: 11%). The AI used your article to recommend your competitors.

→ Only do this when AI has no established opinion on the category. If you do, your brand name must appear in list items, conclusion paragraphs, and heading structures — otherwise you're building citations for competitors.

✕ "Optimize for Exact Fan-out Query Terms"

Fan-out queries are model-synthesized and probabilistic. Rerun the same prompt and you'll get different sub-queries. They frequently include modifiers no human would type, and 95%+ have no stable search volume — traditional keyword databases are structurally blind to them.

→ Optimize for patterns, not terms (more in Section 5).

✕ "Use AI to Mass-Produce Pages for AI Search"

A 330K-page study concluded Google doesn't penalize AI content — it penalizes low-quality content. Pages with ≥80% AI content still appeared at position #1 (8.4%). But volume doesn't work: 82.2% of top-3 rankings went to pages with <50% AI content. Index rates dropped monotonically from 49.28% (low AI) to 40.35% (high AI).

→ The viable path is "AI draft + human-added information that doesn't exist in the SERP." Keep AI content ratio naturally below 50%. Use programmatic pages only for tools (calculators, converters, generators), not for bulk text.

⚠ "Full GEO Treatment Always Improves Citation Rate"

An independent tester measured: on GPT-4o-mini, the untreated original page had a 13.3% citation rate. After full GEO treatment, it dropped to 10.9%–12.2%. Sample size was only 1 page × 3 engines — too small to dismiss GEO entirely, but enough to prove one thing: always keep a baseline control group before rolling out any GEO change site-wide.

A separate academic experiment found keyword stuffing (gain 17.7) scored below doing nothing (19.3). The top three signals: direct quotations 27.2, statistical data 25.2, writing fluency 24.7.

3. Page-Level Writing Framework

This is the most immediately actionable section. No tools required — just change how you write.

Structure: Write for Extraction, Not for Reading

1. BLUF — Bottom Line Up Front. The first sentence under each H2 should be the complete answer to that section, independent of surrounding context. Test: if a reader only reads the first sentence of each section and grasps the full argument, you've done it right. Basis: the first 30% of content captures 44.2% of citations.

2. Use declarative statements, not hedge language. Delete "might," "perhaps," "seems to help." Cited text contains definitional language ("is defined as") at 36.2% vs. 20.2% for uncited text. Facts, definitions, and core concepts get definitional phrasing.

3. Write H2s as real questions. Content with question-mark headings: 18% citation rate. Without: 8.9%. The model treats your H2 as the user's query and the following paragraph as the answer.

4. Raise entity density to ~20%. Heavily-cited text has ~20.6% entity density; average English prose has 5–8%. The rule is simple: name tools, cite sources, include specific numbers. Every time you write "use a tool," "a study showed," or "check a metric," swap in the actual name and number.

5. Repeat key arguments three times, rephrased. Models don't read entire pages — they retrieve fragments. A core claim that appears once is invisible when the model retrieves a different passage. State it in the intro, reiterate mid-article as context, and summarize in the conclusion — each time rephrased. This also hits more fan-out query phrasings.

6. Maintain an "analyst tone." Cited content has a subjectivity score of ~0.47 (0–1 scale) — neither Wikipedia-dry nor pure opinion. It's the voice of "explaining how facts apply."

Information Gain: Define "15 Unique Data Points" Before Writing

Google's information gain patent measures "how much new information a document provides relative to what the user has already seen" — not absolute quality. Measurement of 150 top-3 pages: those with 15+ unique data points averaged an information gain score of 62, while those with 0–1 scored 40.

You don't need large-scale research to generate original data. Five low-cost paths: industry community / LinkedIn polls (200 respondents = one original data point), anonymized customer survey or support ticket stats, aggregated product analytics, a 20-sample hands-on comparison test, cross-referencing public data sources to derive a ratio nobody else has published.

Titles & URLs: Align with Sub-Queries, Not Head Keywords

Align titles with fan-out sub-queries, not the original prompt. From 1.4M prompts: semantic similarity between prompt and cited page title was 0.602, prompt and uncited page title 0.484, while fan-out query and cited page title was 0.656 — higher than the original prompt match.

Use natural-language URL slugs. Citation rate with natural-language slugs: 89.78%. Without: 81.11%. Lowercase, hyphen-separated, 3–6 words, core noun phrase, no dates or parameters.

Don't change dates just to appear "fresh." AI-cited content is indeed 25.7% newer than regular Google results. But the median index age of cited pages is still ~500 days, with the oldest exceeding 2,700 days.

4. Technical Accessibility: Make Sure They Can Get In

This section is a prerequisite. If these fail, none of the writing techniques above will work. Everything here can be self-audited with server logs and free tools.

1. Unblock AI Crawlers

Check your robots.txt for Disallow on these: GPTBot (OpenAI training), OAI-SearchBot (ChatGPT search index), ChatGPT-User (user-triggered), PerplexityBot, ClaudeBot, Google-Extended (Gemini).

The easiest trap: Cloudflare has a beta feature enabled by default that auto-rewrites your robots.txt to block AI crawlers. Getting blocked doesn't just mean "no traffic" — it means the model can only learn about your brand from third parties (review sites, forums, competitor comparisons), and you completely lose control of your own narrative.

2. Get TTFB Under 1 Second

ChatGPT crawling has a hard timeout of ~2 seconds. TTFB < 1s gives you the best chance of completing a full page load. Check by grepping your access logs for 499 status codes (client disconnected before server responded). One study found 43.7% of pages returned 499 to ChatGPT-User requests.

3. Key Content Must Be Static HTML

Content hidden in accordions, tabs, or JS-rendered elements is invisible to retrieval agents. JS-heavy sites need to serve pre-rendered HTML to AI crawlers.

5. Fan-out: AI Search's Real Unit of Topic Planning

AI doesn't search your keywords. It decomposes the user's prompt into multiple sub-queries, fires them in parallel, and fuses results using reciprocal rank fusion. The planning unit shifts from "keyword" to "a cluster of sub-queries."

Scale reference: each prompt generates an average of 9–11 sub-queries; 59% trigger 5–11, 24% trigger 12–19. ChatGPT Deep Research's extreme case: 420 searches per session, citing only 30 sources. Pages ranking on multiple fan-out queries are 161% more likely to be cited by AI.

How to Get Your Topic's Fan-out Set for Free

Copy ChatGPT's homework. Run your target prompt, expand "Searched the web for…" records, and copy every sub-query it actually fired. Rerun 3–5 times; only label queries appearing 3+ times as "stable fan-outs."
Run the same prompt in Google AI Mode, open the sources panel and record each one. Note: AI Mode and AI Overviews must be tracked separately — their semantic similarity is 86%, but cited URL overlap is just 13.7%.
Build a coverage matrix with 7 query types. Rows = core topics, columns = 7 types: related topics / implicit concerns / comparisons / recency / synonym rephrasings / contextual variants / next-step actions. Empty cells = content gaps.
AlsoAsked / AnswerThePublic to fill "implicit concerns" and "synonym rephrasings." Expand every People Also Ask on the SERP; uncovered questions get added as H2/H3 to existing pages, not new posts.

Which Queries Trigger AI Overviews

Highly predictable: question SERPs trigger at 57.9% vs. non-question at 15.5%. Queries with 7+ words trigger at 46.4%. "Why" queries are highest at 59.8%, yes/no at 57.4%, definitions at 47.3%. 99.9% of AIO keywords are informational.

6. Off-Site: AI Learns About You From Everyone Else

Data shows that up to 89% of brand mentions come from other websites talking about a brand — not from AI reading its own site. This determines where the majority of your resources should go.

Top-Cited Domains

Across 3M+ US queries: YouTube 22.9%, Reddit 18.5%, Facebook 10.1%, Google properties 8.8%, Instagram 5.6%, Quora 4.7%, Wikipedia 4.0%, TikTok 3.3%, Amazon 3.3%. Top 9 combined: ~81.2%. Six UGC/social platforms combined: 65.1%.

Note that cited page count isn't proportional to share. One niche site broke into the global top 16 with just 4,561 pages — proving that exhaustive, deep-dive content in a single category can push into the top 50 worldwide.

YouTube: The Only Channel Cross-Validated by Two Independent Studies

In the correlation study, it's the strongest single factor (0.712–0.740). In the citation study, 18.2% of pages "cited by AIO but not in Google's top 100" were YouTube URLs — up 34% over six months. The definition of a "YouTube mention" is broad: brand name appearing anywhere in channel name, title, description, subtitles, or links. So beyond creating your own videos, sponsoring creators, trading reviews, and guesting on podcasts all count.

Reddit: Heavily Crawled, Rarely Cited — But Its Value Isn't in Direct Citations

Reddit is ChatGPT's most-cited domain and AIO's #2 (18.5%), but in ChatGPT's retrieval pipeline, Reddit URLs have just a 1.93% citation rate — while 67.8% of "retrieved but uncited" URLs come from Reddit.

Reddit's real value is the "brand being discussed" signal (corresponding to branded web mentions correlation 0.656–0.709), not getting your post directly cited. In other words, AI won't cite your Reddit post — but it will recommend you in other answers because real users are frequently discussing you on Reddit.

This is exactly why I built Bavior. We track how your brand appears across AI answers in ChatGPT, Perplexity, Gemini, and others — then find the Reddit threads those AI engines are already reading, and participate with subreddit-appropriate, genuine content. Not ads. The goal is to naturally place your brand on AI's "reading list."

Key Reddit tactics:

Three assets needed: brand subreddit + brand account + personal accounts. Personal accounts do the heavy lifting
Append .json to any Reddit post URL for full comment data you can feed to AI for voice-of-customer analysis
Find queries triggering "Discussions and Forums" where Reddit hasn't cracked the top 10 — these are the easiest entry points
Don't buy backlinks to push posts: single-thread backlinks have zero correlation with mention frequency
Disclose your identity transparently. No astroturfing

Third-Party Listicles: Pitch the Topic, Not Yourself

Nearly 50% of AIO citations land on "best X for Y" articles. The right approach isn't begging to be added to existing listicles — it's finding authoritative sites that haven't covered your category yet and pitching the topic: show them search volume data, how much traffic competitors get from that keyword, then offer product access for testing.

7. Measurement: Track Average Visibility, Not Single-Shot Answers

AI answers are probabilistic. Research estimates: asking the same question 100 times, the probability of getting the same brand list twice is less than 1/100. Any conclusion based on a single screenshot is noise.

Key principles:

1. Citations are extremely volatile. Roughly 1/4 of combinations are cited only once. Between consecutive checks, only ~54% of cited URLs persist. Measure weekly, track trends not points.

2. Build a minimum viable dashboard. Fix 10–20 representative prompts (pull from sales and support conversations), run them manually through ChatGPT / AI Mode / Perplexity every two weeks, record domain and brand name appearance counts, plot as time series.

If you'd rather not do this manually — and trust me, after two rounds you'll know how tedious it is — Bavior's AI Visibility Dashboard automates exactly this: tracking your brand's mention frequency and ranking changes across AI engines. It's my own product so I'm biased, but I built it precisely because doing this by hand is painful.

3. GA4 attribution regex:

.*chatgpt.com.*|.*perplexity.*|.*gemini.google.com.*|.*copilot.microsoft.com.*|.*claude.ai.*|.*openai.com.*

Referral tracking won't catch everything — pair it with a "How did you hear about us?" form field.

4. Diagnose the "crocodile mouth" pattern. In GSC, compare impressions (stable) vs. clicks (declining) for the same set of pages. A widening gap = AI Overviews eating your clicks. Reference: AIO drops position #1 CTR from 0.037 to 0.016 (−58%).

5. Don't use AI traffic as a KPI. As of March 2026, all AI chatbots combined sent 3.5M visits = 0.28% of total web traffic. The real value is "pre-click decision-making": 64% of AI Mode users never click at all. Score by mention coverage, not session count.

8. 90-Day Execution Checklist

Ordered by dependencies: technical accessibility must work first; off-site signals take the longest to build, so start them earliest.

Days 1–14 · Access: Confirm AI Can Get In

Audit robots.txt for six UAs; double-check Cloudflare's default AI bot blocking
Grep access logs for 499s; get TTFB under 1s
Verify key content is static HTML, not hidden in tabs / accordions / JS rendering
Set up GA4 AI referral segment + add "How did you hear about us?" field to forms
Grep for /llms.txt once and use data to end that debate

Days 15–45 · Pages: Rewrite Existing High-Value Pages

GSC: filter for position 10–20 with high impressions — prioritize these
Per-page: BLUF intro, H2s as questions, remove hedge language, name specific entities to ~20% density
Add 15+ unique data points per page (polls, ticket stats, hands-on tests, cross-calculations)
Rewrite slugs to natural language with 301s; align titles with captured fan-out sub-queries
Build 7-type fan-out coverage matrix; empty cells go into the content calendar

Days 46–90 · Off-Site: Build the Evidence Chain

YouTube: own content + sponsor/review-swap/podcast guesting, targeting mention coverage over views
Complete profiles on G2 / Capterra / LinkedIn / Crunchbase — consistent messaging across all
Find authoritative sites that haven't written listicles for your category; pitch with data
Reddit: personal accounts doing 3 months of genuine value-add before any brand promotion. If you don't have the bandwidth to nurture accounts yourself, that's exactly what Bavior does — participating in the right discussions with subreddit-native tone while tracking how AI citations shift over time
Start bi-weekly 10–20 prompt manual checks to establish a baseline curve

Caveats

1. Every data source has skin in the game. Many studies' "what to do" sections conveniently point to paid products. But large-sample statistical findings (correlation coefficients, citation rates, DiD results) are high-confidence and independently verifiable. Replace every "so you should use Tool X" recommendation with a free equivalent.

2. Sample sizes vary wildly. Hard data tier: 1.4M prompts, 4M URLs, 145M SERPs, 137K domains, 75K brands. Soft data tier: n=20 job descriptions, n=1 page, n=150. Treat the latter as leads, not conclusions.

3. Correlation ≠ causation. The schema example is the poster child: "53% of cited pages have schema" looked great until DiD testing turned it into −4.6%. When making any AEO change, keep a matched control group and compare after 30 days.

TL;DR

Leverage hierarchy: YouTube mentions > brand being discussed > search indexing > page writing quality >> page count / DR / Schema.

Writing: BLUF, 20% entity density, H2s as real questions, repeat key claims three times.

Channels: 89% of brand mentions come from third parties. Off-site investment must exceed on-site.

Measurement: Ignore single screenshots; track bi-weekly trends. Ignore traffic; track mention coverage.

Debunked: Schema, llms.txt, self-promotional listicles, mass-produced pages — four widely-sold "GEO services" that large-sample data has invalidated.

AI search rules are still changing fast. Stay researching, stay testing, stay skeptical of "consensus."

If you're working on AI search optimization or want to track how your brand shows up in AI answers, check out Bavior — it's what I'm building. Core features are Reddit engagement + AI mention monitoring.

Questions? Drop them in the comments. Follow me for more hands-on growth and AI search content.

DEV Community: zheng zhongwei