AI visibility vs traditional SEO: what marketing teams need to track

The instinct when AI search arrived was to port the SEO scoreboard across: swap Google for ChatGPT, swap position for mention, keep the weekly report. It is a reasonable instinct and it produces bad numbers, because the thing being measured changed category. Not the channel — the measurement itself.

Get this distinction right and the rest of your GEO reporting falls into place. Get it wrong and you will spend a year defending a chart you cannot explain.

The unit of measurement changed

A Google result page is a document. It exists before you look at it, it is substantially the same for everyone in the same location, and you can point at it. Your competitor is at position three whether or not anyone measures that, and if two tools disagree about it, one of them is wrong.

An AI answer is an event. It comes into existence when someone asks, it is assembled fresh each time, and asking again produces a different one. There is no document sitting there with your brand at position three. Two tools reporting different mention rates for the same prompt can both be right, because they sampled different answers.

Rank is observed, mention is estimated

This is the whole thing in one line. A rank is an observation. A mention rate is an estimate of a population parameter from a sample. Different species of number, carrying different obligations.

An observation needs a timestamp. An estimate needs a sample size and a confidence, or it is not a finding — it is a story with a number attached. This is why every credible AI visibility metric arrives as a pair: the rate, and how much you should trust it. A tool that shows you "42%" with no denominator is showing you the output of a calculation it does not want you to inspect.

Metrics that do not carry over

These four have obvious-looking AI analogues. Each analogue is a trap, and each one has quietly ended up in somebody's QBR deck.

• Position. There is no stable slot to occupy. "Mentioned second" in a prose paragraph is not position two — the model may name three tools in one answer and one in the next. Order carries some signal, which is why position is a scored input rather than a headline. It is not a rank, and charting it over time is mostly noise.
• Impressions. In SEO an impression is a logged event. In AI answers there is no impression observable from outside — nobody but the provider knows how many people asked. Any tool showing you AI impression volume has modelled it, and you should ask how before you build a forecast on it.
• Click-through rate. Many AI answers resolve the question without a click. A synthesised answer that recommends you and never links you is a real marketing outcome with no click attached, and a CTR denominator cannot represent it.
• Keyword volume. Prompts are not keywords. They are longer, conversational, frequently multi-constraint ("best CRM for a 12-person agency that needs HubSpot import"), and there is no volume index for them. Prompts should be selected by buyer value, not by a demand number that does not exist for this surface.

What actually carries over

The pendulum swings too far the other way in most GEO commentary, so be clear: a good chunk of SEO work transfers directly, because grounded engines are still reading the web.

• Crawlability and access. If a grounded engine cannot fetch your page, it cannot cite it. The plumbing still matters.
• Structured, extractable claims. Models quote what is easy to lift. A comparison table with explicit numbers is quotable; the same information written as three paragraphs of positioning prose is not.
• Third-party sources. Review sites, roundups, docs and forum threads are disproportionately what grounded engines cite for a category — often more than your own domain. Where SEO taught you to earn links, GEO asks you to earn accurate descriptions.
• Freshness on factual claims. An outdated pricing page or a stale feature list does not just rank badly, it teaches the model something false about you that it will repeat with total confidence.

The honest summary: the technical foundation carries over, the scoreboard does not.

The composite and its weights

Because no single AI metric captures visibility, TrackGeo reports a composite. It is worth showing the weights, because the shape of the formula is an argument about what visibility means.

Visibility Index = mentionRate         x 0.30
                 + positionScore       x 0.20
                 + citationOwnership   x 0.20
                 + sentimentNormalized x 0.15
                 + shareOfAIVoice      x 0.15

Mention rate is the largest single input and still only 30 percent. That is deliberate. Being named is table stakes; being named dismissively, or named while a competitor's source is the one actually cited, is not the same outcome. The formula is fixed and deterministic — every score is arithmetic over stored runs, not a model being asked to rate you out of 100. That distinction matters more than it sounds: a score an LLM invents is subject to exactly the non-determinism you built the tool to escape.

That last row is the point of a composite. A single mention-rate chart would have told you "62.5%, fine". The breakdown tells you the weakest input is citation ownership at 30, which is a specific, ownable problem with a specific fix — get the sources these engines actually cite to describe you accurately — rather than a vague instruction to do better.

Citation is not recommendation

One more conflation to kill, because it has no SEO equivalent and it catches everyone. Being cited and being recommended are different events that can happen independently.

• Cited but not recommended: the engine links your pricing page while recommending a competitor. You are a source, not a candidate. This is common and it looks like success in any tool that only counts citations.
• Recommended but not cited: the engine names you from what it learned in training, with no link. Real visibility, zero referral traffic, and invisible to anything watching your logs.
• Both: the outcome you want, and the only one where the engine's own evidence supports the recommendation it is making.

Track them separately. Collapsing them into one "AI visibility" number destroys the diagnosis — and the two failures have completely different remedies.

What to put in the monthly report

1. 1Mention rate with its denominator and confidence label, on a prompt set you have not edited this month.
2. 2Share of AI voice against a named competitor set — your mentions over your mentions plus theirs. This moves when the market moves, not when you add prompts.
3. 3Citation ownership: of the sources engines cite when answering your category, how many are yours. This is the input most teams can actually move in a quarter.
4. 4Sentiment and any outdated or incorrect claims the engines are repeating about you. A false claim about your pricing is a bug with a fix, not a branding problem.
5. 5Engine spread, because a number that only holds on one engine is a quirk of one model, not a market position.
6. 6Prompt-level movement on the ten prompts that matter commercially — not an average across every prompt you track, which is where real movement goes to die.

The attribution line you should not cross

You can connect GA4 and Search Console and put AI visibility next to sessions and conversions. Do it — the correlation is genuinely useful for deciding where to spend. But hold the line at correlation. AI answers frequently produce no click at all, which means the buyer who read a recommendation and typed your domain in directly is indistinguishable from any other direct visit.

The practical translation of all this: keep your technical SEO discipline, throw out the rank-tracking mental model, and report rates with confidence attached. When a competitor starts taking answers you want, the investigation is a specific one — see "Why competitors appear in AI answers — and how to investigate the gap" (/blog/why-competitors-win-ai-answers). The exact weights and inputs above are documented on /methodology, the per-engine surfaces on /engines, and how the tracked metrics map to the product on /features.