What actually changed
For twenty years, the goal of search optimisation was to be the link someone clicked. That is no longer the only goal. A meaningful and growing share of research now ends inside an AI answer: the user reads a synthesised response, sees three or four brands named, and stops there. No click, no session, no analytics event.
This is not the death of SEO. AI assistants retrieve from the same indexed, crawlable web. But they reward different things: clear entities, extractable facts and third-party corroboration rather than keyword density and link volume alone.
How language models choose what to cite
Language models do not rank pages the way Google does. When answering a question, a retrieval-augmented system typically does three things: it retrieves candidate passages from a search index, it synthesises an answer, and it attributes that answer to sources it judges authoritative and consistent.
From an optimisation perspective, four factors matter most.
- Machinery-accessibility: can the passage be retrieved at all? Blocked crawlers, JavaScript-only content and paywalls remove you from consideration entirely.
- Passage extrability: is there a clean, self-contained sentence or table that answers the question? Long narrative paragraphs with no clear claim give a model nothing to quote.
- Entity consistency: does every source on the web describe you the same way — same services, same locations, same leadership? Contradictions make a model hedge or omit you.
- Third-party corroboration: models weight claims that appear on multiple independent sources. Your own website is one vote; industry publications, directories and review platforms are many.
The practical playbook
You do not need a separate strategy. You need to layer specific changes on top of the SEO work you should already be doing. These are the interventions that have moved citation rates for our clients.
- Publish a canonical fact sheet — one page that states plainly who you are, what you sell, where you operate, who leads the company, and your contact details. Then make every other surface match it verbatim.
- Write answer-first. Begin each section with a single clear sentence that could stand alone as an answer, then expand. Models extract the first claim they find.
- Use tables and lists for comparisons. Structured formats are disproportionately likely to be quoted because they are unambiguous.
- Add an llms.txt file at your domain root summarising your site for AI crawlers, and ensure GPTBot, PerplexityBot and Google-Extended are not blocked in robots.txt unless you have a deliberate reason.
- Deploy complete schema: Organization with sameAs links, Person for named experts, Service and Product for what you sell, FAQ for questions you answer. Make the entity graph unambiguous.
- Earn mentions on the sources models already trust in your sector — industry publications, review platforms, professional bodies and data providers.
- Publish original data. Models cite statistics with attribution far more readily than opinions.
Be the source a model can quote in one sentence, and it will quote you.
How to measure AI visibility
You cannot track this in Google Analytics, because most AI-referred traffic arrives without a referrer. The practical approach is prompt tracking: define a set of 30–150 buyer-intent questions, run them across the major engines on a fixed schedule, and record which brands are named and which sources are cited.
Then measure share of voice against named competitors and track it monthly. The metric that matters is citation rate — the percentage of your tracked prompts where you are named — and it typically moves in 6 to 12 weeks with consistent work.
We also recommend tracking AI-referred sessions in GA4 as a custom channel group using referrer patterns. It undercounts, but it shows direction and identifies which pages AI engines send people to.
What not to do
Two temptations are worth resisting. The first is rewriting everything into short choppy sentences in the hope of being quoted. Readability still matters to humans, and models handle long-form content well when it contains extractable claims.
The second is chasing mentions wherever you can find them. Models weight source quality. Twenty mentions on low-credibility sites count for less than one in a publication your customers actually read.
Key takeaways
- AI assistants retrieve from the indexed web — crawlability and indexation still come first.
- Extractable claims, tables and answer-first sections are quoted far more often than narrative prose.
- Entity consistency across every surface is a prerequisite for being named accurately.
- Measure with prompt tracking and citation rate, not with analytics sessions.
Hamza Tariq
Head of Content & GEO
Runs content strategy and our Generative Engine Optimization practice, including our AI-visibility measurement framework.