A marketing lead pulls up a query her company ranks first for. The blue link is right there at the top. Above it sits an AI Overview that answers the question in four sentences and links to three other sites, and hers is not one of them. Nothing has dropped. Nothing is broken. She is simply not in the box that now sits where attention lands first.
That gap, ranking well yet going uncited, is the real question behind most searches for “AEO” and “GEO.” The honest answer is quieter than the sales pitches make it sound.
It is still SEO, and Google says so plainly #
In 2026 Google published a Search Central guide on optimizing for its generative AI features, and the central message is almost anticlimactic. AI Overviews and AI Mode run on the same core Search ranking and quality systems that already decide who ranks. Google’s current guidance states there are no additional requirements to appear in them, and no special optimizations that unlock them. You do not need a new machine-readable file, an “AI text” file, or special schema markup, and Google notes directly that an LLMS.txt file neither helps nor harms visibility in Search.
The mechanism underneath is retrieval. The system searches its index for relevant, reasonably current pages, then writes a summary and links to a few of them. So the pages it can cite are, by definition, pages it could already find, index, and rank. If a page is indexed, eligible to show with a snippet, and meets the standard technical requirements, it is already eligible to be pulled into an AI answer. Eligible, not guaranteed, the same way ranking has never been guaranteed.
That reframes the work. The lever is not a hidden AI setting, it is the same content quality and technical hygiene that shapes ordinary rankings, aimed at pages that answer real questions cleanly.
Three surfaces, three different behaviors #
“AI search” is not one thing, and treating it as one is where a lot of advice quietly goes wrong. Three surfaces matter right now, and they source their answers from different places.
| Surface | Where it pulls from | What that means for you |
|---|---|---|
| Google AI Overviews and AI Mode | Google’s own index and core ranking, with retrieval over that index | Classic Google SEO and indexability are the entry ticket |
| ChatGPT search | A web index that is not Google’s, having leaned on Microsoft’s Bing | Bing indexation and visibility become part of the picture |
| Perplexity | The live web at answer time, with a lean toward fresh pages | Being crawlable and reasonably current carries more weight |
A page that dominates Google can still be missing from a Perplexity answer, because the index and the retrieval are not the same. The reverse happens too. So the practical target is broad, retrievable visibility across web indexes, not a single trick aimed at one assistant.
Two quick specifics make the split concrete. ChatGPT’s search retrieves from a web index that is not Google’s, and has leaned on Microsoft’s Bing for it, so being indexed and visible in Bing becomes part of the exposure math. Perplexity searches the live web when it answers, which puts a premium on being crawlable and reasonably fresh rather than only well established. Arrangements like these shift over time, but the underlying point holds: different engines look in different places, so no single index owns your visibility.
One more caution sits on top of that split. Most public claims about exactly how each engine picks its sources are third-party observations, not published mechanics. Some are useful directional reads. Many circulate as precise percentages that are hard to verify and tend to shift within a quarter. It helps to treat those numbers as weather, not physics.
What genuinely improves your odds #
Strip away the noise and a short, unglamorous list is left. Google’s own guidance and the shared logic across surfaces keep pointing at the same handful of habits.
- Answer the question early and cleanly. Retrieval can only lift a clear answer. A page that buries its point under three throat-clearing paragraphs hands the model nothing tidy to quote.
- Offer something that is not generic. Google explicitly contrasts a first-hand review, which carries real perspective, against a summary that restates what is already everywhere. The second kind is exactly what a model can produce without you.
- Structure for scanning and extraction. Clear headings, short sections, a direct definition or set of steps where the query asks for one. This overlaps with featured-snippet thinking, and the habit carries over.
- Stay technically eligible. Indexable, crawlable, snippet-eligible, fast enough to be served. None of this is new, and all of it still gates you.
Picture two pages on the same question, say how to clear a specific error code on a common thermostat. One repeats the generic steps every manual already lists. The other shows the actual code, the actual board, the reading that was off, and the fix that held. A system summarizing that query has a reason to reach for the second page, and a reason to name it. Every item on that list also serves a human reader and an ordinary ranking, which is the point Google keeps returning to: build for the audience, and the AI eligibility tends to follow rather than lead.
The myths worth dropping #
A scene from almost any marketing inbox in 2026: a cold pitch promising an “AEO package” that will “guarantee placement in AI Overviews” for a monthly fee. It is worth being clear about what that money cannot buy.
- No special file gets you in. LLMS.txt, AI-specific text files, and bespoke markup do not unlock AI features, by Google’s own statement.
- No secret schema exists for it. Structured data can help Google understand a page, but there is no AI-answer schema that flips a switch.
- Manufactured mentions are a dead end. Chasing inauthentic mentions across the web to look quotable is the kind of tactic that ages badly and can read as manipulation.
- Precise citation-share stats are shaky ground. The “this percent of AI citations come from that source” figures are mostly vendor studies with narrow samples and short shelf lives.
Guarantees are the clearest tell of all. No one controls what a generative system outputs for a given query, so no one can honestly promise a citation for it.
Knowing whether it is working #
Measurement here is genuinely harder than classic SEO, and pretending otherwise helps no one. Google has been folding performance from its AI features into Search Console reporting, and where that is available it is the most direct signal for the Google surfaces. For assistants like ChatGPT and Perplexity, attribution is fuzzier. You might notice referral traffic from them, or catch your own brand being cited by running the same queries yourself, but there is no clean dashboard that counts every mention.
The steadier read is indirect. When the underlying content is genuinely more useful, more first-hand, and better structured than the generic version, it tends to earn both ordinary rankings and AI citations over time. Watching that trend beats chasing any single number.
Frequently asked questions #
Do I need special markup or an LLMS.txt file for AI Overviews?
No. Google states that no special files, markup, or schema are required, and that an LLMS.txt file neither helps nor harms visibility in Search.
Is AEO different from SEO?
For Google’s features, mostly not. They run on the same core ranking and quality systems, so the fundamentals that earn rankings are the same ones that make a page eligible for AI answers.
Does structured data help with AI Overviews?
It can help Google understand and represent a page, and it stays worthwhile for rich results, but Google is explicit that no special schema is required to appear in AI features. Treat it as ordinary good practice, not an AI shortcut.
Why does my top-ranking page get skipped by an AI answer?
Eligibility is not a guarantee, and surfaces retrieve differently. A page can rank on Google yet be absent from a tool that pulls from another index, or from the live web at the moment of the query.
Can an agency guarantee I will appear in AI Overviews?
No provider can. Because no one controls the generated output for a given query, a guaranteed placement is not something anyone can promise honestly.
Where to start #
The short version is almost dull. There is no separate AI back door, just the front door you already maintain. If you want one place to begin, take your most important pages and make each one answer its core question in the first screen, from real experience, on a page clean and fast enough to be served. That work earns ordinary rankings whether or not an AI box ever cites you, which is the safest bet in a field that keeps rewriting its own rules every few months.