Marketing teams have adopted a new vocabulary, and sooner or later it lands on an engineer's desk: "Can you make us show up in ChatGPT?"
It is a fair request, poorly phrased. Nobody can make an AI engine say anything. What engineers can do is control the parts of the system that decide whether a site is reachable, readable and unambiguous. This guide covers those parts, using the definitions in Kiyado's GEO Glossary.
A note before we start: if your team is evaluating generative engine optimization tools, read the definitions first. Most tools measure outcomes. The engineering work below changes inputs.
- Understand the pipeline you are feeding The glossary describes the process in a few terms. Training data is the text a model learned from before release. It is fixed at a knowledge cutoff date and cannot be updated afterwards. Live retrieval is when an AI engine fetches web pages at the moment a question is asked. Retrieval-augmented generation (RAG) is fetching relevant sources first, then writing an answer based on them. Most AI search works this way. Query fan-out is when an engine splits one question into several smaller searches. Google describes AI Mode as issuing many searches on the user's behalf at the same time. Grounding ties an answer to specific retrieved sources, so claims can be traced back. The engineering implication is direct. You cannot change training data after the fact. You can influence what is retrieved live, and how cleanly it can be grounded.
- Control the crawlers deliberately AI companies use their own bots. OpenAI, for example, runs OAI-SearchBot for its search index and GPTBot for collecting training data. They are controlled separately in robots.txt. That separation lets you make a considered decision rather than a blanket one. A pattern some teams use: # Allow OpenAI's search crawler (answers with citations) User-agent: OAI-SearchBot Allow: /
Decide separately on training-data collection
User-agent: GPTBot
Disallow: /internal/
Allow: /docs/
Allow: /blog/
Treat this as an illustration, not a recommendation. Crawler names and behaviour change, so verify each vendor's current documentation before you ship it, and respect every platform's terms of service.
To check what is actually happening, look at your access logs rather than guessing:
Count hits per AI crawler in the last access log
grep -oE "OAI-SearchBot|GPTBot" /var/log/nginx/access.log | sort | uniq -c
Write pages as passages
AI engines extract passages, not whole pages, so each section has to make sense on its own. For a SaaS product, that changes how documentation and feature pages are structured.
A section titled "Overview" that begins "As described above" is a poor passage. A section titled "Does the API support SSO?" that begins "Yes. The API supports SAML 2.0 and OIDC single sign-on on the Business plan" is a good one. (The plan details here are illustrative.)
Practical rules:One question or concept per heading.
Put the answer in the first sentence after the heading.
Repeat the subject's name rather than relying on "it" or "this".
Keep facts such as limits, versions and prices in plain text, not only in images.
Label facts with structured data
Structured data is code added to a page that labels facts explicitly, so machines can read them without guessing. For a software company, organisation and product facts are the obvious starting point:
{
"@context": "https://schema.org",
"@type": "SoftwareApplication",
"name": "Example Analytics",
"applicationCategory": "BusinessApplication",
"operatingSystem": "Web",
"publisher": {
"@type": "Organization",
"name": "Example Ltd",
"url": "https://www.example.com/"
}
}
Validate it, keep it consistent with the visible page, and do not mark up claims the page does not make.
- Make the company a clear entity An entity is a clearly defined "thing" an AI system can recognise, such as your company, a product line or a named expert. Google's Knowledge Graph, its database of entities and how they relate, has shaped search results since 2012. Engineers often own the places where entity facts drift: the docs site, the status page, the app store listing, the developer portal. If the product name is styled three ways across those properties, the entity is less clear than it should be. Pick one form and enforce it.
- Be sceptical of single-file fixes llms.txt is a proposed file that points AI systems to a site's most important content. The glossary is precise about its status: it is a proposal, not a standard, and you should not expect it to fix anything on its own. Adding one is cheap. Treating it as the strategy is a mistake.
- Know what the tools measure When the marketing team buys a tool, it will usually report: Prompt testing results: who gets named, cited and recommended for a set of questions. AI share of voice: how often your brand appears compared with competitors. AI referral traffic: visitors who click a link in an AI answer. The volume is small, but these visitors often have strong buying intent. Those are outcome measures. They tell you whether something changed. The work in sections 2 to 6 is how something changes.
- A practical first sprint If you want to turn this into tickets, a reasonable first sprint looks like this:
- Pull 30 days of access logs and count requests from each AI crawler you allow.
- Review robots.txt with marketing and legal, and document the decision for each crawler.
- Pick the ten most important documentation and feature pages and rewrite their headings as questions, with the answer in the first sentence.
- Add or fix organisation and product structured data on those pages, then validate it.
- Standardise the product name across the docs site, status page and app store listing. None of this requires a new vendor. All of it makes the measurement tools more useful when they arrive. Closing note For an engineering team in London or anywhere else, the honest position is this: you cannot promise placement in an AI answer, but you can remove every technical reason to be left out. That is a worthwhile job, and it starts with shared vocabulary. See how Kiyado works with London technology companies →
Top comments (0)