AI answer visibility: how to measure it with Semalt AI Analytics

For fifteen years, the job of an SEO agency came down to a single question: where do we rank when someone searches for this? That question is still valid. It is no longer sufficient. A growing share of buying decisions now begins and ends inside a generated answer: the user asks, receives a paragraph naming three or four suppliers, and acts on it. They never see a results page. They never click a blue link. And if your business is not named in that paragraph, then as far as that customer is concerned, you do not exist.

The practical difficulty is that this layer of visibility has been invisible to the people it affects. You can see impressions in Search Console and positions in any rank tracker, but you cannot see how often ChatGPT, Gemini or Perplexity recommended your competitor instead of you. That is precisely the gap Semalt AI Analytics was built to close — one of the free modules in the rebuilt platform, and the reason it has become a fixed part of our audit process.

Why AI visibility is a different metric, not a new kind of ranking

It is tempting to treat visibility in AI answers as just another ranking: a list you want to be near the top of. It does not behave that way, and conflating the two leads to expensive decisions.

A classic ranking is deterministic within a narrow margin. Search for "dentist in Seville" today and tomorrow and you will see roughly the same set of results. A generated answer is probabilistic. The same question asked twice can produce two different answers, naming different brands, drawing on different sources. There is no "position three". There is a frequency: across a hundred phrasings of the same intent, how often do you appear?

The second difference is the unit of analysis. In classic SEO you compete for a keyword. In generative search you compete for a question, and questions are long, conversational and carry conditions inside them: "which accountant in Seville handles non-resident landlords", "dental clinic near Los Remedios open on Saturdays that offers financing". Every one of those conditions is a filter your content either satisfies or fails. If your website never states that you open on Saturdays, no model is going to infer it.

The third difference is the uncomfortable one: you can be visible and worse off for it. A model can name you and describe you badly — "a budget option, though less specialised". That is negative visibility, and it has no equivalent in the ranking world, where position one always beats position four. Which is why any serious measurement tool has to read context, not just count mentions.

What the AI Analytics module actually measures

Semalt's approach is refreshingly literal. Instead of selling you an opaque "AI score", it breaks visibility into four dimensions you can work on separately.

1. Queries: the map of questions that matter

The starting point is not your brand, it is the questions. The module lets you research which conversational queries exist around your category and market, and which of them produce answers naming companies like yours. This list is the closest thing we have to keyword research for the generative era, and it usually delivers the first shock: somewhere between 60% and 70% of the questions that matter do not match the keywords you have been optimising for. They are longer, more specific, and loaded with practical constraints — price, hours, coverage area, legal requirements, client type.

2. Mentions and citations: two numbers, not one

A mention is the model naming your brand inside an answer. A citation is the model attributing or linking one of your pages as a source. The gap between the two figures is diagnostic. Heavy mentions with few citations means your brand has recognition but your content is not being treated as a reliable source: the model knows you exist because of what others say about you, not because of what you publish. Heavy citations with few mentions is the reverse — your content feeds answers about your sector without positioning you as the supplier.

3. Sentiment: how you are described, not just whether you appear

Sentiment analysis across answers is, in our experience, the data point that changes client conversations fastest. Seeing in writing that a model characterises your business as "the cheaper alternative" after three years of premium positioning is a useful shock. It is also actionable: that framing comes from somewhere — usually reviews, third-party comparisons, or your own outdated copy.

4. Pages and sources that shape the answer

This is the operational half. The module identifies which URLs — yours and other people's — sit behind the answers. That ends the guesswork. If answers in your category consistently lean on three trade directories and two blogs, you now know exactly where you need to be present and accurately described. It is an influence map, and you work it the way serious link building has always been worked: with judgement and relationships.

The competitive benchmark, where the useful pain lives

None of these metrics means much in isolation. Appearing in 18% of answers in your category could be excellent or catastrophic depending on who is standing next to you. The competitor comparison is what turns the module into a decision tool rather than an attractive dashboard.

The pattern we see most often in established local businesses is this: a technically sound website, good Google positions for its head terms, and a presence in AI answers well below competitors who rank worse than it does. The explanation is almost always the same. Their content was written to win a keyword, not to answer a question. Four-hundred-word service pages with no prices, no conditions, no explicit coverage areas — nothing a model can safely extract and reuse. They rank on accumulated authority, but they supply no citable material.

The opposite diagnosis exists too, and it is equally instructive: small operations with modest domain authority appearing constantly in generated answers because they published honest guides with concrete numbers and clear structure. Generative search rewards specificity in a way the classic algorithm never rewarded it quite so directly.

Fitting AI Analytics into a real working cycle

A tool is only worth the process around it. This is the cycle we run, condensed, and it is reproducible: open your Semalt workspace and connect the domain.

1. Week 1 — BaselineMeasure before touching anything. Define the relevant question set (40 to 60 for a local business, more for ecommerce), record mentions, citations and sentiment, and name the three or four competitors worth comparing against. This first snapshot is non-negotiable: without it, every later improvement is an anecdote rather than a result.
2. Week 2 — Gap analysisCross two lists: questions where a competitor appears and you do not, and questions where nobody answers convincingly. The second list is usually more valuable and almost nobody looks at it — unoccupied gaps, normally in the band of highly specific questions where the competition got bored of writing.
3. Weeks 3 to 8 — Producing citable materialThis is where it is won or lost. Citable material answers the question in the opening paragraph, includes verifiable figures and timeframes, states conditions explicitly (areas served, hours, requirements), is structured in short blocks with headings that mirror the real question, and carries a review date. A model cannot cite what it cannot extract with confidence.
4. Week 9 onwards — Re-measure and correct the framingCompare against the baseline. Citations typically move before mentions — new content enters as a source before the brand gains recognition — and sentiment shifts last, because it depends on third-party material you do not control.

Five mistakes we see on repeat

Measuring the brand only. Asking "what do you know about my company" and calling it research. That measures recognition, not commercial visibility. Customers do not ask about you; they ask about their problem.

Mistaking mention volume for quality. Fifty mentions in answers to irrelevant questions are worth less than three in the question that immediately precedes a purchase.

Writing for the model instead of the reader. Without exception, the texts that perform best in generated answers are texts that are genuinely useful to a person. Padding with synonyms and invented FAQs does not merely fail to help — it introduces noise the model can misquote.

Ignoring third-party sources. If 70% of the answers in your category lean on one trade comparison site where your listing is incomplete, no amount of owned content will compensate.

Treating it as a project rather than continuous measurement. Models update, competitors publish, the underlying corpus shifts. Quarterly measurement is the bare minimum; monthly is sensible.

The local angle: what changes for a city business

Almost everything written about AI visibility assumes a national brand or an international SaaS. A local business is playing a different game, with three specifics worth understanding before spending anything.

Geography enters through the question, not the algorithm. Google knows where you are and filters accordingly without being told. A generative model, absent explicit integration, works with what you type: if the user does not say "in Seville" or "near Triana", the answer will be generic. The consequence is direct — the questions that genuinely matter to you are the ones with the location inside them, and there are far fewer of them than your search volume data suggests. Forty well-chosen questions cover practically the entire real market of a clinic, an accountancy practice or a workshop in the city.

Competitive density is low, and that opportunity has an expiry date. In local categories we still routinely find clearly commercial questions where no city business appears convincingly: the model falls back on generic criteria or cites national directories. Whoever publishes specific, verifiable material first occupies that gap for a modest content investment. In two years that window will be shut, exactly as the local SEO window shut around 2012.

The sources feeding local answers are predictable. Google Business Profile listings, reviews, local press, professional associations, trade bodies and a handful of directories. When you inspect which URLs sustain answers in your category, the list rarely exceeds fifteen domains. That short list is your reputation and presence workplan: complete profiles, consistent data across all of them, and descriptions that say what you want repeated about you.

Translated into budget: a local business does not need an AI strategy. It needs thirty honest, well-structured, properly localised pages and a serious review of a dozen external profiles. The work is affordable; knowing where to aim is the hard part, and that requires data.

What makes the Semalt approach different

Two reasons this module matters to us beyond its novelty.

First, it does not live in isolation. The same account carries Search Console analytics, rank tracking in Google SERP, and fast indexing. This matters more than it sounds. When you publish a guide designed to be cited, you want to push its indexing, watch it enter organic positions, and check whether it starts feeding generated answers — all against the same clock. Holding those three data points in three separate tools, each with its own definition of "last 30 days", is the most efficient way to conclude nothing at all.

Second, the commercial model. The Semalt analytics layer is free with no time limit; the paid side is managed campaigns. That inverts the industry's usual arrangement, where you pay before you are allowed to see the diagnosis. For a business that does not yet know whether it has an AI visibility problem, being able to measure it before committing budget is exactly the right order of operations.

Frequently asked questions

Conclusion: measure first, argue later

Generative search has produced far more opinion than data over the past two years. It will kill SEO; it will change nothing; citations are all that matter; citations are irrelevant. Almost none of those claims arrive with a measurement of a specific business in a specific market attached.

Our recommendation to any company asking these questions is the least glamorous one available: build a baseline this month. Define your forty questions, measure where you appear, measure where your competitors appear, and store that figure. In a quarter you will hold something almost nobody in your sector has — your own time series, from which you can decide on evidence instead of reacting to headlines.

To start with the free layer, open your workspace, connect the domain, and spend an afternoon building your question list. It is comfortably the highest-return hour of work available to your business's visibility right now. If you would rather we set it up and interpret it with you, we now include it in our initial audit — just get in touch.