The three layers of AI search — and the AEO strategy for each.

This is the layer most people meet first, because it sits on top of the search engines they already use. When Google generates an AI Overview, it isn't reasoning from memory — it's reading the live web, right now, through the same crawl infrastructure that powers regular search. Bing Copilot works the same way.

That means the live-parse layer rewards machine-readable infrastructure. The engines read your structured data, your Google Business Profile, your knowledge panel, and the consistency of your name, address, and phone number across every directory that mentions you. When those signals agree, the engine can treat your business as a confirmed entity and name you with confidence. When they conflict — an old address on one directory, a different phone number on another — it quietly drops you, because naming the wrong business is worse than naming no business.

The strategy here is unglamorous and concrete: clean schema on every page, a claimed and complete Google Business Profile, a knowledge panel that agrees with your website, and name-address-phone consistency across every directory that mentions you.

This is also the fastest layer to move. Because it reads the live web, fixes show up as soon as the next crawl picks them up. In our engagements, infrastructure cleanup typically shows results in two to six weeks. If you do nothing else, do this — it's the foundation the other two layers build on.

The second layer is where most of the lead-driving queries get answered. Perplexity, ChatGPT with browsing, and Gemini with browsing all work from a query-time web index: when a buyer asks a question, the engine retrieves a handful of sources it trusts, reads them, and composes an answer with citations.

The operative phrase is sources it trusts. These engines don't read the whole internet for every answer — they reach for a short list of authorities per category. And that list is remarkably stable: a category review platform (G2 for software, Clutch for agencies, Yelp for local services), the directories your industry actually uses, and a handful of city-level editorial outlets. If those sources mention you for a service, you have a real chance of being named. If they don't, your own website — however good — is usually not enough, because the engine treats third-party mentions as evidence and self-description as marketing.

Take your highest-intent query — say, "best [your service] in [your city]" — and run it through Perplexity. Look at the citations under the answer. That short list of sources is your actual battlefield. Not the answer. The citations.

The strategy for this layer is citation building: getting your business named, reviewed, and quoted in the specific publications each engine reaches for in your category. That means a deep, active profile on the review platform your category trusts, with service-level detail. It means editorial outreach — pitching the writers who actually cover your category, not press-release spam. It means showing up in the round-ups and "best of" lists that engines love to quote, because a ranked list maps perfectly onto the three-names-per-answer format.

Timeline: weeks, not days. Citations have to be published, then crawled, then weighted. Editorial placements take longer — pitching cycles are real — but they compound, because a single trade-press mention gets retrieved across thousands of future queries.

The third layer is the slowest and the most misunderstood. When ChatGPT answers without browsing — which is most of the time for casual users — it answers from training data: a frozen snapshot of the web, months old. Claude works the same way. No fix you ship today appears in this layer until the next model is trained and released.

This frustrates people, but it cuts both ways. The training-data layer is slowest to enter and hardest to lose. A competitor can't displace you between retraining cycles, no matter what they publish this quarter. Visibility here is the closest thing AI search has to a moat.

What gets a business into training data? Volume and consistency over time. The model learns entities the way it learns everything else — from repetition across independent sources. A business that appears in editorial coverage, directories, review platforms, and local press, described the same way everywhere, becomes a stable entity the model can recall. A business with three mentions and two conflicting descriptions becomes noise, or worse, a hallucination.

The strategy is the same citation and editorial work as layer two, sustained — plus entity discipline: same business name, same specialties, same descriptions, everywhere, for long enough that the next training run can't miss you. Expect 60 to 90 days minimum before a retraining cycle can even pick the work up. We tell clients to treat this layer as a deferred asset: you build it now, it pays later, and it keeps paying.

Being named is half the job. What the AI says about you is the other half, and it's the part almost nobody audits. We regularly find engines that name a business but describe the wrong specialties, an old location, or a team member who left two years ago. One hallucinated detail in an AI answer reads as fact to the buyer — and you'll never see it, because nobody sends you a screenshot of their ChatGPT session.

This is why every engagement we run starts with a query map: 80+ real buyer queries across all five major engines, recorded with what each engine actually said. Not just hit or miss — what it claimed, what it got wrong, and which source the error traces back to. Correcting a hallucination is source work: find the stale or wrong page the engine is leaning on, fix or displace it, and re-test until the answer changes.

A growing pile of AEO advice is either useless or harmful. For the record: stuffing your pages with "best in [city]" self-praise does nothing, because engines discount self-description by design. AI-generated content farms get filtered first and distrusted second. Hidden-text tricks asking models to recommend you get caught and burn credibility you can't easily rebuild. And nobody can "submit your site to ChatGPT" — there is no such pipe, whatever the cold email says.

The honest version of AEO is structural: make your business legible to machines, get the sources machines trust to vouch for you, and keep the record accurate. That's it. Everything else is decoration.

Knowing the layers is one thing; sequencing the work is another. We run every engagement in the same order, because the dependencies only run one way.

Weeks one and two are pure infrastructure: schema, knowledge panel, NAP cleanup, review architecture. This is the live-parse layer, and it goes first for two reasons — it moves fastest, and everything downstream depends on it. A trade-press citation pointing at a business whose own entity data is a mess is a wasted citation; the engine can't connect the praise to a confirmed identity.

Weeks three through eight shift to the web-index layer: review-platform depth, directory citations, the first round of editorial pitches. Placements land on their own schedule, but the pipeline has to start early because everything here has a publication lag before engines can retrieve it.

Beyond that, the same citation work — sustained and kept consistent — becomes the training-data play by default. You don't run a separate "layer three campaign." You keep the record deep and uniform until the next retraining cycle picks it up. The layers aren't three projects; they're one project with three different settling times.

Each layer needs its own measurement. We compress it into a weekly visibility score — high-intent query coverage, position in the answer set, authoritative citations, and sentiment accuracy — and we pair it with branded search lift from Google Search Console, because when an AI names you, people Google you within a day or three to verify. Most businesses start somewhere between 15 and 35. Infrastructure work moves the live-parse layer first; citations move the index layer next; the training-data layer follows a cycle later.

Three layers, three timelines, one discipline. Businesses that understand the difference stop asking "why am I not on ChatGPT yet?" and start asking the better question: "which layer is this week's work aimed at?"