AI search optimisation checklist

You will see this called AEO, GEO, LLM SEO or AI search optimisation. The acronym does not matter and it will change again. What matters is a narrower question:

When someone asks an assistant a question you could answer, are you the source it cites?

Eighty per cent of this is SEO you already know

Let me be honest about that before selling you a checklist. Crawlability, genuinely unique substance, a named accountable author, accurate structured data — these were already the things that worked, and they matter more here, not less.

What is genuinely different is worth knowing precisely:

Aspect	Classic search	AI search
What wins	The best page for a query, judged partly on links	The most extractable, best-sourced passage that answers the specific question
Unit of competition	The page, against nine other pages	The passage, against passages from several sites blended into one answer
What a win looks like	A ranking position and a click	A citation, and often no click at all
Measurement	Impressions, position and clicks in Search Console	Manual prompt checks and referral traffic. There is no Search Console for this yet
What does not transfer	Keyword density, exact-match phrasing, thin pages at volume	All of it works against you — filler dilutes the passages you want extracted
What transfers completely	Crawlability, unique substance, a named accountable author, accurate structured data	The same four, weighted more heavily

The last two rows are the important ones. Thin pages at volume actively hurt you here — filler dilutes the passages you want extracted — while the fundamentals transfer completely.

The checklist

5 groups · 26 checks

Can the assistants reach you at all? 5

• robots.txt allows the crawlers you want: OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended
• You have made a deliberate decision on GPTBot and Google-Extended rather than leaving the default
• No WAF or bot rule is challenging those user agents — check your firewall logs, not just robots.txt
• Content is in the HTML rather than requiring JavaScript to render
• Pages return 200 quickly; slow responses get abandoned by crawlers as well as people

Is the answer extractable? 6

• Each page answers its main question in the first 40–80 words, before any preamble
• One page answers one question — a page covering five topics gets cited for none
• Headings are real questions or plain statements, not clever labels
• Facts sit in sentences and tables, not inside images or charts
• Tables have real headers so a row can be read without the surrounding prose
• Lists are short and parallel, so an extract does not lose the meaning

Is it worth citing? 5

• The page contains something not available elsewhere — original data, a real file, a tested verdict
• Numbers carry a source and a date
• The author is named, with checkable credentials on a real profile
• The page states when it was last updated, and that date is true
• You say what you do not know, or where the approach fails. Hedged pages get cited more, not less

Machine-readable signals 5

• Structured data describes what is visible on the page and validates
• A single Person or Organization entity, with the same @id used sitewide
• sameAs links to the profiles that corroborate the entity
• Canonical, title and description are unique and accurate on every page
• No FAQ markup invented for content that is not on the page

Measuring it 5

• You have a list of the prompts a buyer would actually type
• You check those prompts across at least two assistants on a fixed schedule
• You record whether you were cited, mentioned, or absent — and which URL was used
• You watch referral traffic from chatgpt.com, perplexity.ai and similar in analytics
• You accept that impressions here are largely unmeasurable, and do not invent a number

Deciding which crawlers to allow

This is the part that is genuinely new, and it is routinely confused. Blocking a training crawler does not stop you being cited in search answers. They are separate decisions.

User agent	Who runs it	What it does	Advice
OAI-SearchBot	OpenAI	Surfacing pages in ChatGPT search results	Allow. This is how you get cited.
GPTBot	OpenAI	Collecting data used for model training	Owner’s decision. Blocking it does not stop you being cited by OAI-SearchBot.
PerplexityBot	Perplexity	Indexing for Perplexity answers, which cite sources prominently	Allow.
ClaudeBot	Anthropic	Crawling for Claude	Allow unless you have a reason not to.
Google-Extended	Google	Controls Gemini training use — NOT Search indexing	Blocking it does not affect Google Search rankings. Decide separately.
Bingbot	Microsoft	Bing index, which also feeds Copilot	Allow. Blocking it costs you Copilot visibility too.

Writing a passage that survives extraction

The unit of competition is the passage, not the page. Four things follow from that:

Answer the question in the first 40–80 words, before context or preamble. An assistant extracting an answer takes the passage that contains one, and a page that warms up for three paragraphs offers nothing to take.

One page, one question. A page covering five topics gets cited for none of them, because no single passage in it fully answers anything.

Put facts in sentences and tables, not in images. A number inside a chart is invisible. A number in a table row with a real header can be read on its own.

Say what you do not know. This one surprises people: hedged, caveated pages get cited more, not less. “This works for X but fails at Y” is more useful to a system trying to answer accurately than a confident claim it cannot verify.

Being mentioned somewhere else

You will be told that getting discussed on Reddit, LinkedIn or YouTube is the fastest route to being cited. There is something real here and it is worth being careful about.

What appears to be true: assistants draw on community and discussion sources heavily, so a recommendation of you somewhere else can surface where your own page does not. Being the answer on a page you do not own is still being the answer.

What is not a tactic: posting about yourself. Communities like Reddit are unusually good at spotting it, and the outcome is a ban and a deleted account rather than a citation. Paying for placement in “best X” listicles has the same problem in a slower way — those pages get discounted as the pattern becomes obvious.

What actually works, and is slow: being genuinely useful in places where your buyers already are, under your own name, without linking to yourself most of the time. That is a reputation strategy with an SEO side-effect, not an SEO tactic — and it is worth being honest that it takes months and cannot be outsourced cheaply.

Measuring it, honestly

There is no Search Console for AI search. Anyone offering you a precise “AI visibility score” is offering a modelled estimate.

What works is unglamorous: a list of prompts a buyer would actually type, checked across two or three assistants on a fixed schedule, recorded. The tracker is where that goes.

Column	What to put in it
Prompt	The words a buyer would actually type, not a keyword.
Intent	Comparison, how-to, recommendation, definition. Dropdown.
Assistant	ChatGPT, Perplexity, Claude, Google AI Overview. Dropdown.
Date checked	Answers change week to week, so an undated check is worthless.
Cited?	Yes / Mentioned / No. Dropdown. "Mentioned" means named without a link.
Your URL used	Which page it pulled. Often not the one you would expect.
Competitors cited	Who was cited instead. The most useful column on the sheet.
Gap	What their cited page has that yours does not.
Action	What you changed. Then re-check on a stated date.

The competitors cited column is the most useful thing on the sheet. It tells you what a citable page looks like for your topic, which is far more actionable than your own citation rate.

This site is the worked example

Everything on this page is applied here, which is the fairest way to judge it. Every resource opens with a 40–80 word direct answer that the build fails without. robots.txt names the assistant crawlers explicitly. There is one Person entity with the same @id sitewide and sameAs links to profiles that corroborate it. Facts sit in tables with real headers, and the pages say where the approach fails.

Whether it works is measurable by you: ask an assistant something this site should be able to answer, and see who it cites. How the resources are made sets out the rest of the method.

Common mistakes