Is Your Business Getting Recommended, or Just Indexed? A Framework for Telling the Difference

Three questions hiding inside one

Owners usually ask a single question, "am I showing up in AI search," and expect a single answer. The evidence says that question contains three distinct ones, and the answer to each can differ. First, is your page in the index or corpus at all. Second, when a buyer asks a question you should own, does the engine retrieve and cite your page in the live answer. Third, of the businesses the answer names, are you the one it recommends. A framework that collapses these into one number will mislead, because the mechanisms that govern each are different.

This is not a semantic distinction. It maps directly onto how modern answer engines are built. Generative answers are produced by retrieval-augmented generation, an architecture that pairs a language model with a separate index queried at the moment of the question, so that the answer is assembled from passages pulled at inference time rather than recited from memory. Presence in that index is a precondition, not an outcome. The steps that run after retrieval, ranking, generation, and citation, decide whether presence becomes a recommendation.

Being indexed was never the finish line

The idea that the index is a substrate rather than a destination predates generative AI by more than a decade. When Google introduced its Knowledge Graph in 2012, it framed the shift as indexing "things, not strings," building a store of entities and facts that search could reason over rather than a flat list of matched documents. The visible ranked list has always been a thin interface over a deeper representation of who and what exists.

Retrieval-augmented generation extended that logic. The founding work showed that pairing a parametric model with a non-parametric index produced more specific and factual output and could update its knowledge without retraining, precisely because the index is consulted fresh each time. What follows is a plain consequence: your page being in the corpus tells you almost nothing about whether it will be chosen from the corpus. Indexation is table stakes. The interesting question begins after it.

Ranking and citation have come apart

The most consequential split for an owner is between classic rank and AI citation. Industry analysis of Google's own technical disclosures describes AI Overviews assembling a query-specific custom corpus of candidate passages, ranking them, and then running a separate citation pass after the answer text has been written. If that description holds, two things follow that break the old mental model. The passages used to compose the answer and the pages credited beneath it are not guaranteed to be the same set, and a page ranking well outside the top organic results can still be cited while a page ranking first is left out.

We tier this claim as emerging rather than established, because it rests on secondary technical analysis of Google's patents and documentation rather than a single peer-reviewed primary source, and it should be read as directionally supported, not settled. But the direction matters. It means "we rank number one" is not evidence that you are cited, and "we are not cited" is not evidence that you rank poorly. Ranking and citation are now two different races, which is the structural reason a serious diagnostic measures classic search and AI answers as separate pillars rather than one blended score.

A citation is not a click, and often not a faithful read

Suppose you clear both bars: you are indexed, and you are cited. Two independent bodies of evidence show that a citation is still weaker than owners assume it to be.

The four states, from absent to recommended

Put the evidence together and a business, for any given buyer question, sits in one of four states. The framework is diagnostic: you locate the state, then you know what the actual problem is.

• Absent. Your page is not in the index or corpus for this topic. The answer cannot cite what it does not hold. The problem is indexation and technical foundation.
• Indexed but not surfaced. Your page is in the corpus, but it is not retrieved or cited in the live answer to the questions that matter. The problem is relevance, structure, and corroboration, not mere presence.
• Surfaced but not recommended. Your business is cited among several, but the answer names a competitor as the one to choose, or your citation is one nobody acts on. The problem is competitive standing and the strength of your entity relative to the businesses named ahead of you.
• Recommended. The answer names your business as the one to choose across the buyer questions you should own. This is the only state that reliably moves customers, and it has to be measured question by question, engine by engine, because it does not hold uniformly.

How to tell which state you are in

No engine publishes whether it recommends you, so direct observation is the only way to find out. The procedure is simple to state and disciplined to run. Assemble a panel of the real questions your buyers ask, phrased the way they phrase them, not the keywords you wish they used. Run that panel across each engine that matters to your market, for example ChatGPT, Google AI Overviews, Perplexity, Gemini, and Copilot. For each question, record which of the four states you are in: absent, indexed but not surfaced, surfaced but not recommended, or recommended. Note who is named instead of you when you are not.

Two cautions matter here. Generative answers are not deterministic, so the same question can return different sources on different runs, which means a single check is anecdote and a repeated sample is evidence. And the answer surface barely forwards clicks, as the Pew data shows, so you cannot infer your AI-answer state from your analytics traffic. It has to be observed on the answer surface itself. The output of this procedure is not a rank. It is an appearance rate with a date and an engine attached, reported as a range rather than a single confident number, precisely because the surface is volatile and undocumented.

What actually moves a business from indexed to recommended

The framework would be incomplete without the evidence on what changes the state, and here the record is unusually clear. The founding, peer-reviewed study of generative engine optimization tested content interventions in a controlled benchmark of roughly ten thousand queries. Adding citations to credible sources, including direct quotations, and replacing vague claims with specific statistics produced a 30 to 40 percent relative lift on a position-adjusted visibility metric, and citing authoritative sources was consistently the single strongest lever. The object being optimized is no longer a rank; it is whether a source is worth quoting.

A large-scale 2025 follow-up study found that AI answer engines skew toward earned, third-party media over brand-owned and social content at a systematically higher rate than classic Google does. We tier that finding as emerging, since it is a single large study not yet independently replicated, but it points the same way as the GEO benchmark: the signals that move a business from indexed to recommended are substance and outside corroboration, not volume of thin pages. None of this is a guarantee. What it does is tell you which levers have measured effects, so the work targets the right state.

What does not move it: two common misconceptions

A diagnostic is only as good as the false leads it rules out. Two of the most common in this space fail on the primary evidence, and a framework that respects the reader has to say so.

The first is the llms.txt file, promoted as a way to hand AI systems a clean summary and thereby control visibility. An analysis of 137,210 domains found that 97 percent of valid llms.txt files received zero requests in a single month of 2026, and Google has stated publicly that the file is not used for search. Placing a file at your root is not a path to being recommended. The second is the belief that E-E-A-T, experience, expertise, authoritativeness, and trust, is a score you optimize a page into. Google's own documentation is explicit that E-E-A-T is a human-rater evaluation framework used to assess the algorithm's output quality, not a machine-computed ranking factor a page can directly target. Much commercial AI-visibility advice conflates a rater heuristic with an optimization dial. The diagnostic value of naming these is simple: it keeps budget aimed at the levers with evidence, and away from the ones with only marketing.