How AI engines decide who to recommend
When ChatGPT names three providers and stops, it isn’t consulting a league table. It’s assembling a sentence from three layers of evidence — and each layer is something you can work on.
Ask an assistant for “the best HVAC company near me” and the reply arrives in seconds: three or four names, each with a plausible reason. No ranking algorithm published, no ad auction visible. What actually happened is that the engine assembled a sentence from three layers — what it already knew, what it just looked up, and what it felt safe to assert. Understanding those layers is most of understanding AEO.
Layer one: what the model already knows
Engines are trained on a snapshot of the public web — directories, review platforms, articles, forum threads. Businesses that were widely and consistently described before that snapshot get a head start: the model “knows” they exist, what they do, and roughly how they’re regarded. Thinly documented businesses are, to the model, barely rumors — and rumors don’t get recommended.
Layer two: what it looks up at answer time
For local and commercial questions, most engines now retrieve live sources before answering — maps data, review sites, “best of” articles, industry directories. The reply is grounded in whatever those sources agree on. This is the layer you can move fastest: the citations attached to real answers in your category are a literal to-do list of where your business needs to be present and well described.
Which sources those are varies by category, and the variation matters. Answers about lawyers lean on legal directories and bar listings; answers about home services lean on maps profiles and review platforms; B2B categories pull industry lists and comparison articles. There is no universal checklist — which is why a real measurement reports the sources the engines actually cited for your questions, not a generic audit.
The special case of Google’s AI surfaces
AI Overviews and AI Mode sit directly on top of Google’s index and business data, so the signals classic local SEO cultivates — the business profile, review volume and recency, category labels, proximity — carry extra weight there. An engine like Perplexity shows you its citations openly; Google shows you fewer receipts but leans on infrastructure you may already rank in. Same discipline, different mix — and another reason to measure the surfaces separately rather than assume one result speaks for all.
Layer three: what it feels safe to say
Engines are tuned to avoid asserting doubtful things. A business whose sources disagree — two addresses, three category labels, a website that doesn’t load — is a risky sentence, and risky sentences get replaced with safer competitors. Consistency isn’t cosmetic; it is what makes you speakable.
The signals that make a business safe to say are mundane and checkable: the same name, address, phone and category everywhere; a live website that states in plain words what you do and where; reviews that are recent as well as numerous; descriptions that agree about your specialty; and third-party pages that mention your service and city together. None of it is glamorous. All of it is legible to a machine deciding whether to put your name in a sentence it will be judged on.
Why the same question returns different names
Even with every layer in your favor, generation is sampled: the same prompt, re-run, produces different lists. The largest public test of this behavior found the shuffle is the norm, not the glitch.
SparkToro / Gumshoe.ai (Rand Fishkin, Patrick O’Donnell), January 2026
What stays stable is not any single answer but each brand’s frequency of appearing across many runs. That is the quantity worth managing — and the only one worth believing in a sales pitch. We wrote up the full argument in why we run every prompt ten times.
What engines can’t do
Three inabilities shape the whole game. Engines can’t inspect your business directly — no site visit, no phone call — so they proxy everything through what sources say. They can’t take payment for organic mentions, which is why every “guaranteed AI placement” pitch is selling something the seller doesn’t control. And they can’t carry your reputation for you: an engine that named you yesterday re-derives the answer from evidence today. The evidence is the asset; the mention is just its shadow.
This is also why every page you publish now has two audiences. A human skims it deciding whether to call; a machine reads it deciding whether you’re safe to recommend. Plain sentences that state who you serve, where, and at what price range work for both. Clever copy that says nothing works for neither.
What this means for your business
- You can’t script the answer, but you can stack the evidence — training data and live retrieval both read from the same public sources.
- The citations in real answers tell you exactly which sources matter in your category; work those first.
- Consistency across sources is a ranking factor in the plainest sense: it decides whether you’re safe to name.
- Judge progress in frequencies over repeated runs — never in one screenshot, yours or a vendor’s.
The practical playbook is in how to get recommended by ChatGPT, and the standing scoreboard is the leaderboard — measured monthly across six engines with the method public. If you want your own number, the free scan is the honest place to start.