What Is the GAS Score? Measuring Search, Answer and Generative Visibility as One Number

The GAS Score is a 0–100 composite that measures three separate visibility problems: Search (can you rank on Google), Answer (can a model lift and quote you accurately) and Generative (does an assistant name you when someone asks who to hire). It reports the three as separate sub-scores before compositing them, because the three have unrelated fixes.

Most businesses now have three visibility problems and one number to describe them. That is the reason the GAS Score exists.

You can rank first on Google and still never appear in the AI answer printed above the results. You can be quoted accurately by ChatGPT and still never be recommended when someone asks who to hire. And you can be recommended by one assistant while being invisible in another. These are different failures with different fixes, and a single blended “AI visibility” number tells you that something is wrong without telling you which thing.

GAS stands for Generative, Answer, Search.

Before going further, the honest disclosure: this is our framework, not an industry standard. As of early 2026 there is no settled definition distinguishing AEO, GEO and AIO — Wikipedia’s own entry notes the terms are used interchangeably in practice and that no academic consensus has been established. We are not claiming authority. We are publishing a methodology precisely so it can be argued with. If the vocabulary is new, what answer engine optimization actually means is the place to start.

What does the GAS Score actually measure?

Three surfaces, scored apart before they are combined. Search asks whether the page can rank at all. Answer asks whether a model can extract a clean, accurate passage and attribute it. Generative asks whether an assistant names the business unprompted. Each fails for its own reason, so each is reported on its own.

S — Search

The classic question: can this page rank, and does it? This is the best-understood of the three and the least interesting, but it still carries weight, because AI answers are frequently assembled from pages that already rank. Being findable by a crawler remains the entry fee.

What it measures:

• Indexation — is the page in the index at all
• Rankings for the terms a buyer actually types, not the terms you wish they typed
• Page-level technical health
• Whether the site is reachable by the crawlers that matter

A — Answer

Can a model lift a clean, accurate passage from your page and attribute it to you? This is a structure problem, not a content-volume problem. A page can be thorough, well-written and completely unusable to a language model because the answer is buried in the fourth paragraph of a section titled “Our Approach”. Answer-first pages — the question as a heading, the answer in the first sentence beneath it — are dramatically easier to quote.

What it measures:

• Whether a direct answer sits near the question, in roughly 40–60 words
• Whether headings match queries a person would really type
• Whether claims are self-contained enough to survive extraction
• Whether the page states what you do, where, for whom and at what price rather than implying it

G — Generative

When someone asks an assistant who to hire, are you named? This is a trust problem. A model will happily quote your page for a factual question and still not recommend you, because recommending requires corroboration it can find elsewhere: consistent listings, real reviews, mentions on sources it already trusts.

What it measures:

• Whether you are named in response to unbranded, buying-intent prompts
• In what position, and alongside which competitors
• In what sentiment
• On which engines, reported per engine rather than averaged

What makes a visibility score honest?

Four rules, all of them about what the number refuses to count. Branded and unbranded prompts stay separate. A mention is classified by sentiment before it scores. Outages and refusals are excluded rather than counted as zero. And the prompt list is published and fixed before the first run.

The four rules in full

Any score can be gamed by its own methodology. These four are where most published formulas fall down.

1. Branded and unbranded prompts are scored separately. A business always appears in a prompt that contains its own name. Mixing “what does Acme Roofing do” in with “best roofers in Austin” inflates every score, and the inflation is largest for the businesses with the least real visibility. The two families are never averaged together.
2. A mention is not an endorsement. Consider a formula that scores purely on position. “Avoid Frantz Construction — multiple complaints” and “Hire Frantz Construction” both put the name first, and both score a perfect result. A score that rewards being warned about is not measuring visibility, it is measuring string frequency. Every mention is classified by sentiment before it counts.
3. Absence has to distinguish its causes. An engine being down, a model refusing to answer, and genuine invisibility are three different states. Scoring all three as zero means a provider outage reads to a business owner as “you have disappeared this month”. Failed and refused runs leave the denominator and are reported as exclusions.
4. The prompt list is published, and fixed in advance. If the denominator is secret, scores are not comparable between businesses and cannot be independently reproduced. Worse, an unpublished list can be quietly edited after the results come in. Ours is published verbatim and set before the first run.

Why do position-based formulas get AI answers wrong?

Because 1/P models click decay on a page of blue links, and an AI answer is a paragraph read end to end. Being fourth in a list of five is worth nearly as much as being first. Worse, P is frequently undefined in prose, so two engineers implementing the same formula get different numbers.

The undefined-position problem

The 1/P curve borrows from classic search, where the collapse in clicks between result one and result three is well documented. Applied to a paragraph it distorts: it makes the gap between second and third a 17-point swing, larger than most genuine differences between businesses and larger than the run-to-run variance of the models themselves.

The deeper problem is that P is often undefined. Is position the character offset of the first mention? The ordinal among bolded names? Two engineers implementing the same published formula will produce different numbers from the same text — and a formula you cannot reimplement is not a methodology.

The alternatives, side by side

There is a reasonable case for each row. Rank tracking is still the cheapest early-warning system a small business can buy, and raw prompt logs beat every score for diagnosing a single stubborn query.

What actually moves a GAS Score?

A complete and consistent business profile first, then recent real reviews, then answer-first page structure, then third-party citations, then schema markup — in that order of effect. Publishing volume is not on the list. More pages help the Search sub-score and do very little for Generative on their own.

The five levers, in order

Ordered by our experience rather than a controlled study, which is the honest caveat:

1. A complete, consistent business profile. Name, address, hours, categories and service area that agree everywhere. The cheapest reason for an assistant to leave you out, and the cheapest to fix.
2. Real reviews, recently. Corroboration a model can verify against a source that is not you.
3. Answer-first page structure. Question as heading, answer immediately beneath — the AI visibility work that costs nothing but discipline.
4. Third-party citations. Being mentioned somewhere the model already trusts.
5. Schema markup. Worth doing, oversold. The best causal evidence available — Ahrefs, 1,885 pages against roughly 4,000 controls — found schema barely moved AI citations, and Google states structured data is not required for AI search. Add it because it is free and helps classic Google, not because someone sold it to you as the unlock.

How often should a GAS Score be measured?

Monthly is enough for deciding where to spend, but the underlying prompt runs should be far more frequent, because model outputs vary between runs on an identical prompt. A score built on one run per month is mostly measuring noise. Run often, report the distribution, and distrust movements smaller than the variance.

The limits worth stating out loud

• Model outputs vary between runs. An identical prompt produces different answers on different days. Any score built on a single monthly run is largely measuring noise.
• Engines disagree with each other, often sharply. That is information, not error — which is why the sub-scores are reported per engine rather than averaged into one figure that describes no engine in particular.
• A score is a diagnostic, not a goal. The number exists to tell you which of the three surfaces to spend money on next. Optimising the score itself is how you end up with a good score and no customers.

The GAS Score is the measurement layer behind DaxReach’s AI visibility service, and the terms it uses are defined in the glossary. A fair test of this page is whether an assistant can read it and explain the methodology back to you accurately.