AI search engine optimization is the work of getting your brand named and cited inside AI-generated answers, rather than ranked in a list of links. The surfaces that matter are ChatGPT, Perplexity, Gemini and Google AI Overviews, and the unit of success changes from position to citation. Most of the technical groundwork is shared with SEO, but the measurement is not, and that is where teams get it wrong.
Disclosure: GetIntel sells AI visibility tracking, so we have a commercial interest in this topic. Every statistic below is attributed to a source you can check, and where the evidence is weak, llms.txt in particular. This guide says so rather than selling around it.
In this article
- What is AI search engine optimization?
- How is it different from SEO?
- What actually gets a brand cited?
- Does llms.txt help?
- How do you measure AI visibility?
- What do teams get wrong?
What is AI search engine optimization?
It is optimising to be the source an engine reaches for when it writes an answer. A buyer no longer scans ten links and picks one; they ask a question and get a composed reply naming two or three products. Either you are named or you are not.
That reframes the goal. Ranking is a proxy for attention when the user chooses from a list. When a model composes the answer, the only outcomes are being recommended, being cited as a source, or being absent, and those are three different things worth measuring separately.
How is it different from SEO?
The groundwork overlaps; the measurement does not. Crawlable HTML, clean structure and schema markup help both, Google's own AI optimization guide is explicit that there is no separate technical checklist for AI surfaces.
What changes is what you count. SEO measures rank and clicks. AI search measures whether you are named, which sources the engine cited, and where in the answer you appeared. Those do not move together: a page can rank first and never be cited, and a page that ranks nowhere can be quoted because it answers one specific question cleanly.
| SEO | AI search | |
|---|---|---|
| Unit of success | a ranking position | being named or cited in the answer |
| What you count | rank, clicks, impressions | mentions, citation share, position in the answer |
| Who decides | the ranking algorithm, from a list | the model, composing one reply |
| Where the work lands | your own pages | your pages and the third-party sources engines quote |
Two findings are worth holding onto. Roughly 21% of keywords now trigger an AI Overview, and about 97% of AI Overview results cite a page already ranking in the top 20, both figures from SEOProfy's AI SEO statistics roundup, which credits the first to Ahrefs and the second to SeoClarity. The practical reading: conventional ranking still buys you the ticket, it just no longer decides the outcome.
Separately, around 60% of searches end without a click, per Semrush data collected in SlateHQ's statistics roundup. That figure is about search generally rather than AI Overviews specifically. It is often quoted as though it were AI-caused, and it is not.
What actually gets a brand cited?
Being the clearest answer to a specific question, on a page a crawler can read, from a domain the engine already has reason to trust. There is no fourth lever anybody has demonstrated.
In practice that means:
- Answer the question in the first sentence. Models lift self-contained passages. A paragraph that opens with context and buries the answer in sentence four is harder to quote.
- Structure for extraction. Headings phrased as questions, tables for comparisons, FAQ blocks. Our own gap data suggests FAQ-structured content is cited more often by AI Overviews than the same content in prose, though that is an internal figure we have not published.
- Say where numbers came from. An unsourced statistic is a liability once an engine starts attributing it to you.
- Be present where engines look. Third-party sources, Reddit, G2, review roundups, carry a large share of category citations, often more than a vendor's own site.
Search Engine Land's guide to generative optimization covers the tactical layer in more depth than fits here.
Does llms.txt help?
There is no confirmed evidence that it does, and you should be sceptical of anyone selling it as the fix.
SE Ranking's study across roughly 300,000 domains found no measurable citation effect, reported by Search Engine Journal, which notes the file was found on only about 10% of domains crawled — and Google has said it does not use the file. It costs nothing to publish and does no harm, so publish one if you like — but treat it as a small, unproven signal rather than a lever. We build llms.txt generation into our own product and still would not claim it moves citations on its own.
The honest version of the Foundation layer is unglamorous: make sure AI crawlers are not blocked in robots.txt, that your content renders without JavaScript, and that your schema is valid. Those are verifiable. llms.txt is not, yet.
How do you measure AI visibility?
Put a fixed set of real buyer questions to each engine on a schedule, and record three things per answer: whether you were named, which domains were cited, and where you appeared in the response.
Three rules make the numbers usable:
- Keep the prompt set fixed. Rewriting a prompt changes the answer, so a changed set makes your own history incomparable.
- Measure each engine separately. Engines cite very different sources from one another; a blended score hides which surface actually moved.
- Record the date. These systems change underneath you, and an undated figure is not a measurement.
The failure mode is a single composite number that goes up and down with nothing actionable behind it. Presence, citation share and position answer different questions, and collapsing them loses exactly the detail that tells you what to do next.
What do teams get wrong?
Treating AI search as one surface. It is not. Engines overlap far less than the category assumes — in our own 100-prompt runs, most cited domains appear on a single engine rather than across several. Optimising for one tells you little about the others.
Chasing a score instead of a question. "Our AI visibility is 34" is not actionable. "We are absent from six of the ten questions our buyers ask, and a competitor is cited in five of them" is.
Publishing more instead of publishing clearer. Volume was an SEO strategy. Citation rewards the page that answers one question better than anything else available, which is a different production problem.
Quoting statistics without checking them. This category recycles numbers badly — figures get attributed to studies that do not contain them, and general-search data gets relabelled as AI-specific. If you cannot find a number in its cited source, do not repeat it.
If you would rather not assemble the measurement by hand, that is what GetIntel does: it tracks the questions your buyers ask across ChatGPT, Perplexity, Gemini and Google AI Overviews — with Claude on the Growth plan — shows which sources each engine pulled, and drafts the fix when a competitor is cited instead of you. If you want the measured version of any of this, we publish it: how many domains each engine cites, and how far apart Google's three AI surfaces are. You can start free, or read how our AI visibility tracking works.