The State of AI Visibility: A Baseline Reading (Volume 1 of the Surface Report)

Why AI visibility needs a baseline no one has yet published

There is no shortage of commentary about AI search. There is a shortage of trustworthy measurement. No generative engine publishes a full account of which businesses it names, how often, or why, and the numbers that circulate are largely vendor-produced, methodologically opaque, or drawn from a single account. For an individual owner the practical question is narrower and unanswered by any of it: when a buyer asks an engine who does this kind of work, how often does my business appear in the answer, and how does that compare with the businesses that appear instead of me?

A baseline is the precondition for answering that with evidence. Without a reference distribution built from real businesses, any single reading is a number without a scale, and any claim of improvement is an assertion rather than a measurement. The Surface Report exists to build that reference distribution and to publish it on a fixed cadence, so that a later reading can be compared against an earlier one and against a peer set, rather than against a marketing figure.

The strongest independent evidence available today establishes the problem the index addresses, not its solution. It shows that the behavior of AI answers is measurably different from the behavior of the classic results page, which is precisely why the old yardsticks do not carry over and a new baseline has to be constructed rather than borrowed.

What the established evidence already shows, and what it does not

The Surface Report is not a leap of faith. It rests on a body of primary evidence that is independent, methodologically transparent, and in several cases peer-reviewed. That evidence is strong enough to justify measuring AI visibility as its own discipline. It is not, by itself, a baseline of who wins, which is the gap the index is built to close.

AI search statistics are abundant and mostly unreliable

The reason a careful index is needed rather than another dashboard is that the existing supply of AI search statistics is noisy and, in places, actively misleading. Two documented examples set the standard the Surface Report has to clear.

The first is llms.txt, the proposed root-level file promoted as the way to hand AI systems a clean summary of a site. An analysis of 137,210 domains found that 97 percent of valid llms.txt files received zero requests in May 2026, and Google has stated publicly that the file is not used for search. A tactic marketed as a control lever is, on the evidence, read by almost nothing. The second is E-E-A-T. Google's own documentation describes Experience, Expertise, Authoritativeness, and Trust as a framework its human raters use to evaluate the algorithm, and Google has repeatedly clarified that it is not itself a scored ranking factor a page can optimize into a document. A large share of commercial advice conflates the rater heuristic with a machine-readable target.

The pattern is consistent: the tooling layer sold as how you control AI visibility is younger, leakier, and less standardized than the marketing suggests. An index that wants to be trusted has to be visibly stricter than the field it reports on, which is why the method is published before any number is.

What the Surface Report will measure: the four pillars of the Machine-Readiness Score

The unit of the index is the Machine-Readiness Score, a 0 to 100 reading composed of four pillars, held separately rather than blended into a single rank. The pillars are classic search, the local map pack, AI answers, and reputation. They are kept distinct because the evidence says they behave differently and move for different reasons.

The AI-answer pillar is the one the Surface Report is built around, and it is defined against the evidence above. It measures citation and mention presence across the major generative engines, explicitly not organic traffic, because the Pew data shows organic traffic understates this surface. A business is scored on how often it is named when a real buyer question is asked, and on the quality and consistency of the sources the engine reaches for when it answers.

The remaining three pillars anchor the reading in the surfaces that still carry volume and still decide local trade. Classic search and the local map pack remain where a large share of intent lands, and reputation, the corroborating record of reviews and third-party mention, feeds both the classic vote and the AI citation. Separating them matters: a single blended number would hide exactly the decoupling that makes this era worth measuring.

Generative engine optimization and the decoupling of rank from citation

The intellectual case for measuring classic search and AI answers as separate pillars is that ranking and citation are now measurably different jobs. The generative engine optimization literature identifies content levers that lift citation, and those levers are not identical to the ones that lift a classic rank. A 2025 large-scale controlled follow-up went further and found that AI answer engines are systematically biased toward earned, third-party media over brand-owned and social content, a sharp contrast with classic Google, which the same study found sources more evenly. That finding is a single study and not yet replicated, so the Surface Report treats it as emerging rather than settled, but it points the same direction as the founding paper: the sources an engine trusts to answer a question are not the same set that a classic ranking rewards.

The direct implication for the index is that a business can rank on page one and still be absent from the answer written above it, and can be cited from a page that sits far outside the top organic results. A number that averaged the two together would report a comfortable middle that describes neither. Measuring them apart is the only way to tell an owner which of the two races they are actually losing.

Share of voice ai: how the baseline will be sampled

The AI-answer reading is a share-of-answer measurement. For a given business, the method samples a panel of the real questions its buyers ask, phrased the way buyers phrase them, and puts each question to the major engines. It records how often the business is named, alongside which competitors, and drawing on which sources. Repeated on a fixed cadence and paraphrased across wordings, that produces a share-of-answer figure and a movement over time, rather than a one-off screenshot.

Aggregated across many businesses under a consistent protocol, those individual readings become the Visibility Corpus, the reference distribution the Surface Report publishes from. The corpus is the reason Volume 1 is a method and not a headline: a cross-business baseline is only trustworthy once the corpus holds enough entities, across enough verticals, that a reported figure is a reading and not an anecdote. Until that threshold is reached, publishing a baseline would be exactly the kind of thinly-sourced number this index is meant to replace.

The protocol is stated in advance on purpose. Pre-registering what is sampled, how questions are selected, and how presence is scored is what lets the first published reading be judged against a standard that was set before the result was known, which is the difference between a measurement and a marketing figure.

The rules the Surface Report holds itself to

An index is only as credible as the rules it refuses to break under pressure. The Surface Report commits to a short list, in writing, ahead of the first release.

• No fabricated or naked number. Every figure carries its source and its method, and no metric is reported without the sampling behind it.
• Method before numbers. The protocol is published first, so a later reading can be checked against a standard set in advance rather than one written to fit the result.
• Evidence is tiered by strength. Established, emerging, and contested findings are labeled as such, and a single unreplicated study is never dressed up as settled fact.
• No guaranteed outcome. The index reports where a business stands and how it moves. It does not promise a ranking, a citation, or a position.
• The corpus must earn the baseline. No cross-business figure is published until the Visibility Corpus holds enough entities to report a reading rather than an anecdote.

What Volume 1 does and does not claim

To be unambiguous: this installment publishes no cross-client Machine-Readiness Score baseline, because a trustworthy one does not exist yet. The Visibility Corpus is still accruing, and a number reported before it reaches usable scale would violate the first rule on the list above. What Volume 1 delivers is the standard the first real reading will be held to, grounded in independent, primary evidence for why AI visibility has to be measured as its own thing.

The reading itself follows when the data supports it. In the meantime the starting point is not a published industry average, it is a measured read of a single business, the same read that, aggregated under one protocol, becomes the corpus this report will one day publish from.