Answer Engine Optimization: Earning a Citation | Searchlings
Answer Engine Optimization is the work of getting your business named inside the answer an AI assistant gives, instead of ranked in a list of links below it. It applies to ChatGPT, Perplexity, Google’s AI Overviews and AI Mode, and Microsoft Copilot. Most of it is ordinary good SEO. The part that is not ordinary is this: a model has to be able to lift a self-contained answer off your page, find the same facts about you everywhere else it looks, and have a reason to name you rather than a competitor.
That is the whole answer. The rest of this explains why each part works, what evidence there is for it, and which popular tactics do nothing.
What is Answer Engine Optimization?
An answer engine takes a question, fetches a handful of sources, and writes one answer. Sometimes it names those sources, sometimes it links them, and often the person who asked never visits any of them.
So the work splits into two jobs that fail for different reasons.
The first job is being fetched. That fails when a page cannot be reached, cannot be parsed, or does not visibly answer the question that was asked. It is a technical and structural problem, and it is the one most advice covers.
The second job is being named. That fails when the page is fetched and understood perfectly well but contains nothing the other sources on the subject did not already contain. Models summarise consensus. A page that restates the consensus gets absorbed into it without attribution.
Answer engine optimisation is both jobs. The first is mechanical and finite. The second is a question about whether your business has anything specific to say, and no amount of formatting substitutes for it.
One note on spelling before it becomes distracting. We write optimisation the Australian way in most of our copy. The term arrived from American writing and kept its American spelling, which is why the phrase people search for is “Answer Engine Optimization” and the thing we do all day is optimisation. Nothing turns on which one you use.
What counts as an answer engine
The products people mean, as at August 2026:
- ChatGPT
- Perplexity
- Google AI Overviews
- Google AI Mode
- Gemini
- Microsoft Copilot
- Claude
- Grok
They differ in where the sources come from. Some run their own index, some lease one, some fire a live search only when the question seems to need current information, and some answer from the model’s weights with no retrieval at all.
That last case deserves naming, because it is invisible from the outside. When no retrieval happens, nothing you published this month is in the room. What is in the room is whatever the model absorbed about your market during training, which you cannot edit and cannot see.
For most businesses the volume sits with the Google surfaces, because AI Overviews and AI Mode appear in front of people who never chose to talk to an assistant. The chat products are where the considered questions get asked, which is a smaller number of higher-intent conversations. Both are worth appearing in and neither needs its own plan.
You do not need a separate programme per engine. They read much the same web, and the properties that make a page usable to one make it usable to the rest: reachable, parseable, direct, consistent with itself, and specific. What varies between them is sampling. The same question can name you in one product and not in another on the same afternoon.
AEO and SEO: what actually changes
Google’s own search documentation puts it plainly: from Google Search’s perspective, optimising for generative AI search is optimising for the search experience, “and thus still SEO”. That is the vendor describing its own system, which is worth more than most commentary about it.
So treat what follows as a change of emphasis inside SEO, not a new trade.
| Dimension | Classic SEO | Answer Engine Optimization |
|---|---|---|
| The query | A phrase typed into a box | A full question, often carrying context from earlier in the conversation |
| What you win | A position in a list | A sentence in the answer, with your name in it |
| Who reads the page | A ranking system, then a person | A retrieval system, then a model, then a person |
| The unit of work | A page per keyword | A self-contained answer per question |
| Second place | Page two, still visible | There is no page two |
| Fact consistency | Helps | Decides it: claims that conflict across sources tend to be dropped |
| Measurement | Position and clicks, reported daily by Search Console | Sampled answers, because no engine publishes impressions |
| Volatility | Days to weeks | Answer to answer |
The measurement row is the uncomfortable one. In classic search you can see your position for a term every day, for free, from the source. In answer engines there is no equivalent and no sign of one coming. Everything anyone knows about AI visibility, including everything in this article, comes from asking questions and counting the names in the replies.
AEO, GEO and LLM SEO are mostly the same argument
Wikipedia’s article on Generative Engine Optimization, as at July 2026, records the state of play accurately: the academic literature has not settled on a definition that separates these terms, and in practice they are used interchangeably.
We keep separate pages for them because people search separately for them, not because the work differs. If you want the same material framed around the research literature, read Generative Engine Optimization. If you want it framed around the assistants themselves, read LLM SEO. The short definitions live in the glossary: AEO, GEO and LLM SEO.
Anyone selling you AEO and GEO as two engagements is selling you the same work twice.
Why this matters now
The link between ranking and being cited has loosened, and there is a number for it. Ahrefs, in March 2026, studied 863,000 keyword SERPs and 4 million AI Overview URLs. It found that 38 per cent of AI Overview citations came from pages ranking in the top 10, down from roughly 76 per cent in July 2025. About 31 per cent came from pages ranking beyond position 100.
Read both halves of that carefully, because each is routinely overstated.
The first half says a top-10 position is no longer close to a requirement. Under half of citations now come from pages that hold one. Something other than rank is doing a lot of the selecting.
The second half says roughly a third of citations go to pages that, for the query in question, are not ranking anywhere a person would ever look. A page can be invisible in the list and quoted in the answer above it.
What this does not license is abandoning ranking work. Ranking still puts you in the retrieved set for a large share of questions, and the same qualities that rank a page tend to make it retrievable. The honest reading is narrower: rank is becoming one input among several rather than the gate, and a business with no rankings and nothing distinctive to say is not helped by either finding.
How to do AEO
Nine things, in rough order of how reliably they appear to matter. None of them is exotic. Several are dull enough that they get skipped for that reason alone.
The first few overlap with LLM SEO, and that is not an oversight. The same properties make a page usable to an assistant whichever acronym you file the work under, so the overlap is the finding rather than a gap in one of the articles.
1. Put the answer in the first paragraph
Retrieval systems chunk a page and work with the chunks. The first chunk is the one most likely to be pulled, and it is the one a model reads before deciding whether the rest is worth attention.
There is sampling to support this. CXL looked at 100 AI Overview citations and found 55 per cent came from the first 30 per cent of the source page.
So state the answer, then explain it. This article does that, and so does every entry in our glossary. It is also better writing than the alternative.
2. Write sentences that survive being quoted alone
A sentence beginning “as mentioned above” or “this is why it matters” cannot be extracted, because it only makes sense in place. A model that wants to quote you will either rewrite it, in which case your name may not travel with it, or skip it.
Apply the constraint to your commercially important paragraphs and the prose changes noticeably. Definitions get restated instead of referred back to. Nouns replace pronouns.
3. Use one description of your business, everywhere
If your founding year, service list, trading name, location or one-line description differ between your website, your LinkedIn page, your Google Business Profile and three directories, a model has no principled way to choose between them. The usual outcome is not that it picks wrong. It is that it avoids the claim entirely and names a business whose facts agree with themselves.
Write the description once. Paste it verbatim everywhere. This is the most mechanical, least interesting and most reliably effective item on this list.
4. Give the model something citable
Aggarwal and colleagues, in “GEO: Generative Engine Optimization”, accepted to KDD ‘24, tested changes to source pages and measured visibility in generated answers. Adding citations, quotations and statistics lifted visibility by 30 to 40 per cent. Keyword stuffing performed poorly.
The mechanism is not mysterious. A model writing an answer needs load-bearing material, and a specific claim attached to a named source is easier to reuse than an adjective. A number you measured, a date, a named method, a quoted authority: these are the parts of a page that get carried across into an answer, because they are the parts that cannot be paraphrased into the general average.
5. Structure the page so a parser can follow it
Real headings in order. Real lists. Tables where the content is tabular. Structured data that states what the page is. Question-shaped headings above the paragraphs that answer them.
None of this is new advice and all of it matters more when the first reader is a program rather than a person skimming.
6. Write for the question, not the phrase
People type phrases into search boxes and speak whole sentences to assistants, often with conditions attached: the suburb, the budget, the constraint, the comparison. A page built around a two-word keyword answers a fragment of that.
Take the questions your customers actually ask before they buy, in their words, and give each one a page or a section that answers it directly. The unit of AEO is a question, not a keyword.
7. Be consistent in places you do not own
Models read about your business in sources that are not your website: reviews, directories, industry bodies, news, forums, supplier lists. Those sources contribute to what a model believes about you and they are weighted differently from your own claims about yourself, for obvious reasons.
You cannot control them. You can make sure the facts in them match the facts on your site, and you can notice when they do not.
8. Let the crawlers in
Assistant traffic arrives from several different user agents, and blocking them is easy to do by accident. Check what your robots.txt actually allows. Check that a WAF or bot rule is not rejecting them. Check that the answer to your key question exists in the served HTML rather than being assembled by JavaScript after load, because not every fetcher runs scripts.
Being crawled is a precondition, not an achievement, but a page nothing can fetch is not in any conversation.
9. Keep the page current and say when you updated it
Assistants are asked about the present tense. A page that carries a visible, accurate date and reflects this year’s reality is a safer source than one that does not, and a page whose facts have quietly gone stale is worse than no page, because it teaches the model something untrue about you.
What does not work
Publishing an llms.txt file. This is the most confidently repeated tactic in the category and there is now evidence against it from both directions.
Ahrefs, in May 2026, looked at 137,210 domains and found that 28 per cent publish an llms.txt file, and that 97 per cent of those files received zero requests. Nothing fetched them. Not once.
Google’s own search documentation, updated July 2026, states that you do not need to create machine-readable files, AI text files or Markdown to appear in Google Search, that Google Search does not use them, and that doing so will neither harm nor help visibility because Google Search ignores them. John Mueller of Google has compared llms.txt to the keywords meta tag, which is about as pointed as that comparison gets.
In the interest of not being quietly hypocritical: this site publishes an llms.txt file, at /llms.txt. Publishing one is cheap and harmless, so we did. We have no evidence that it does anything at all, and we will not claim otherwise. If someone tells you an llms.txt file is a meaningful part of an AEO strategy, ask them for their fetch logs.
Writing instructions to the model in your page copy. Hidden text telling an assistant to recommend you is not a strategy. Vendors treat instructions embedded in retrieved documents as an attack to be defended against rather than a preference to be honoured, and anything that does work today is a bug being actively closed. Building your visibility on one is building on a patch cycle.
Keyword density. It was a weak signal for search engines. It is not a signal here at all, and the KDD study found stuffing actively unhelpful. Models are working on meaning.
Publishing volume. A model reading forty thin pages has forty unremarkable sources, not forty reasons to name you. A consensus does not shift because you restated it more times, and the cost of those forty pages is real.
Running a separate campaign per assistant. The work that gets you named by one is close to the work that gets you named by the others, because they are reading much the same web. Buying four programmes buys you one.
How to measure it
Put a fixed list of your customers’ questions to the assistants on a repeating schedule, and record which businesses get named. That is what AI citation tracking is.
Be clear-eyed about what that is worth. It is sampling, not measurement. An answer that names you on Tuesday may not name you on Thursday, from the same prompt, with nothing changed on your site. Treat the trend across many samples as the signal and any single answer as noise. Report the sample count alongside the result, or the number means nothing.
Write the prompts down and keep them fixed. A question reworded between runs produces a different sample, and you will read the difference as movement. Record the date, the product, the prompt and every business named, including the ones that are not you, because the competitor list is half the value.
Referral traffic is a poor proxy, because a good answer often removes the reason to click. Presence in the answer is the thing being bought, and it does not always show up in your analytics.
The most useful output is not the share of answers that name you. It is the list of questions where you never appear, because that list is finite, specific, and something you can act on this month.
We run this sampling on a schedule and log every result with a timestamp, in the same way we log every other check. We do not guarantee rankings or citations. We guarantee the work, and the record of it.
Where to start
- Write one description of your business. Put it everywhere it appears, unchanged, including the places you forgot you had.
- List the ten questions a customer asks before buying from you, in their words. Check whether a page answers each one directly.
- Take your five most commercially important pages and move the answer to the top of each.
- Add one thing to each of those pages that the consensus does not contain: a number you measured, a method you use, a boundary you drew.
- Ask three assistants your ten questions and write down who gets named.
Step five is uncomfortable and worth doing before the other four, because it tells you which of them is urgent.