Getting Cited in AI Overviews: What Actually Works
by Arbaz
Answers now appear above links, and a share of the people who would have clicked through read the summary and stop. That is a real shift and it deserves a serious response.
It has also produced a great deal of nonsense, because almost nothing in this field can be verified by the buyer. What follows separates what is observable from what is being asserted.
The strongest correlation is the least exciting one
Pages that already rank conventionally are the ones that get pulled into generated answers.
Google's AI Overviews draw on pages Google already indexes and evaluates. Assistants that retrieve live results are, broadly, using conventional search under the surface. So the page that wins the generated answer is usually a page that was already competitive for the query.
This is why "SEO is dead, do AEO instead" is a sales position rather than an observation. The foundational work is the same work: get crawled, get indexed, rank, and be clear.
What appears to make a page extractable
Observable patterns, stated as patterns and not as guarantees.
It answers one question directly and early
A question answered in the first two sentences under a heading matching that question is far easier to lift than the same answer in paragraph nine.It contains specifics a model does not already hold
Numbers, prices, dates, named entities. A page saying a retainer costs QAR 2,900 offers something quotable. A page saying pricing is competitive offers nothing to quote.It is structurally parseable
Real headings, genuine lists and tables, proper FAQ markup. Decorative formatting that looks like structure is not structure.The entity is unambiguous
Who you are, where you operate, what you sell, stated consistently across the site and reinforced in structured data, so nothing has to be inferred.It is current
Generated answers skew toward recent sources on anything time-sensitive, and stale pages lose ground faster here than in conventional results.
None of that is exotic. It is the discipline that makes a page good for a human reader, applied with more care about extraction.
The three surfaces behave differently
Google AI Overviews and AI Mode
draw on the index. Conventional ranking is the strongest input.ChatGPT
answers from training data plus retrieval. Retrieval favours clear, well-structured, unambiguous pages.Perplexity
was built around citation and is the most transparent about its sources, which makes it the most useful surface for testing, because you can see what it used.
There is no single lever that moves all three. Work that improves all three is mostly conventional work done carefully.
What cannot be measured, said plainly
This is the part that gets left out.
There is no rank tracker for generated answers in the way there is for blue links. Answers vary by user, by phrasing, by session and by time. Ask the same question twice and you can get two answers citing different sources.
So a dashboard reporting "AI visibility: 73" is presenting an estimate as a measurement. The number moves and nobody can tell you exactly why.
What is defensible is sampled testing: a fixed set of questions your buyers actually ask, run on a schedule across the three surfaces, recording whether you appeared and what was cited instead. That carries real uncertainty and should be reported as a sample. It is genuinely useful for direction and it is not a ranking report.
Three things being sold that are not real
"LLM submission services."
There is no submission process. You cannot submit a site to ChatGPT the way you submit a sitemap to Google. This is being sold in this region and it does not exist.Guaranteed citations
Nobody controls what a generated answer includes. A guarantee here is less verifiable than a page-one guarantee, which makes it easier to sell and no more honest.AI-written content at volume
Publishing fifty generated articles a month produces fifty pages containing nothing a model did not already know, which is precisely the wrong thing for extraction.
Who should spend on this now
Worth doing
if you sell something with a long consideration period where buyers research before contacting anyone: professional services, healthcare, property, business setup, B2B. Those are the questions people increasingly ask an assistant instead of a search box.Worth doing
if your category has genuinely poor public information. Qatar has several. Freehold property rules for foreign buyers is the clearest example: the rules are confusing, what exists online is stale or contradictory, and whoever explains it properly gets cited because the alternatives are weak.Wait
if your customers search on a phone with intent to buy immediately. Nobody researches an emergency callout through a chat interface. That budget belongs in the map pack.Wait
if your conventional foundations are broken. Fixing indexing and rankings improves AI visibility as a side effect. Doing it in the other order does not work.
A test you can run this afternoon
Write down the five questions a customer asks before buying from you. Ask each one in Google, ChatGPT and Perplexity. Record whether you appear, and if not, who does and what the cited page looks like.
That costs an hour and tells you more than any dashboard will. Usually the cited page is one that answered the question directly, early, with a specific number in it.
The AI search page goes further into the three surfaces and what is honestly measurable. This work sits inside the Authority plan rather than being sold separately, because on its own it would be a service with no defensible measurement attached to it.