AI visibility agency or AI visibility tool: which one you need

What a visibility number costs to buy

Start with the price of the thing you are actually considering. In August 2026 we audited six AEO and GEO providers, and the self-serve end of that set publishes subscription tiers of $99, $299 and from $499 a month, with execution sold on top as per-unit credits. Another publishes a from-$6,000 monthly price for a platform your own team operates with a strategist in the loop. There are also free properties in the market: one of them publishes category-level AI visibility rankings at no cost, and it runs a category covering AEO and GEO providers, which means our own market is being scored daily by a competitor's free product. All of those figures are what those providers publish about themselves, captured on 2026-08-15.

So the raw number is close to a commodity. What a dashboard sells you is a percentage, refreshed on a schedule, next to a competitor list. What it cannot sell you is the judgement about whether that percentage means anything, and it cannot sell you the work that changes it.

The rest of this essay is about the first of those, because the second is obvious. If nobody on your side is going to rewrite a page, earn a mention or fix an indexation problem in response to the chart, the subscription is a monthly payment for a feeling. That is a description of what software is rather than a criticism of it.

The number is only as good as its denominator

Every AI visibility figure is a ratio. The numerator is the part nobody controls: how often an engine named you. The denominator is chosen, and that is where the meaning lives.

Two choices produce very different numbers from identical captures. A per-run rate asks in what fraction of executed prompt runs you were named. It converges as you add runs, and it falls when you are absent, which is what a reader expects a visibility figure to do. A union asks on how many prompts you were named at least once across all runs. A union can only rise. Every extra run is another chance for a hit and no chance for a miss, so it climbs toward one hundred percent as a function of how long the tool has been switched on.

We shipped exactly that error in our own product, and it is the reason we are confident others have too. Until 2026-06-02 our leaderboard computed the union across all runs and labelled it visibility. One internal account read 75 percent on that basis where the corrected per-run rate was 38.2 percent. Nobody gamed anything; the metric simply answered a different question from the one its label claimed. We renamed the union to reach, made the per-run rate the headline, published both definitions with the sample size beside them, and the regression that proves the fix is that the rate now holds steady as runs accumulate while the old union climbed with every one.

The prompt set is the other half of the denominator, and it moves the number just as hard. Drop the prompts you never win and share rises. Add prompts containing your own brand name and it rises again. Two prompt sets of the same size are not the same basis: one client's first and second sets were both 55 prompts and shared only 17 of them, so a month-over-month line drawn across that change would have been a rewrite rather than a trend. Reading current classification flags against historical runs once produced a 48-point swing in one client's favour on a single day, which is why every figure we compute is pinned to the prompt-set cohort that was live when the runs executed.

None of this is exotic. It is the ordinary reason two vendors report different numbers for the same brand, and before comparing two figures a buyer should compare the four things underneath them: the prompt set, the run count, the engines, and whether the figure is a rate or a union. Usually one of the four explains the entire gap.

Variance, and why a screenshot is not a reading

Engines do not return the same answer twice. That is not a defect in the measurement, it is the nature of the thing being measured, and it has three consequences a buyer should insist on.

First, presence is a rate. AI Overview presence for the same query varies by roughly 15 percentage points between runs in our own measurement, so the only defensible report is a rate over repeated passes with the number of passes printed next to it. A vendor screenshot showing an overview, or showing none, is one sample of a noisy process.

Second, absence cannot be proved by one instrument. Our capture proves an overview was present. An empty return is consistent with a genuine absence, an unexpanded asynchronous block and a rate limit, and those are indistinguishable in the output. Anyone reporting you are absent from AI answers, on the strength of one tool's empty result, is reporting a property of their tool.

Third, a figure below a certain sample size should not travel at all. Our rule is that a citation figure has to have run at least three times before it leaves the building. That rule is expensive and we apply it to ourselves: every client citation share we currently hold is a single-sample reading with roughly 25 points of margin, which is why there is no client citation percentage anywhere on this site. What we can show instead is measurement with its denominator attached. For 603 Basement Solutions, who have cleared us to use their name, a 488-point local grid across 8 service keywords and 61 towns showed them holding a local pack position at 10 points, or 2.0 percent, on 2026-08-09, while one competitor held 38.1 percent of the same grid. That is a diagnosis you can check the arithmetic on.

One score is the wrong shape for the problem

Most visibility products resolve to a single headline number, sometimes a branded score out of one hundred. That shape is convenient and it destroys the information you are paying for.

Citation happens in two stages with opposite fixes. Entering the pool of sources an engine will retrieve for a question is governed by authority: rank, host class, corroboration. Winning the selection once you are in that pool is governed by structure: whether the answering sentence exists as a clean passage, whether the page is the shape the engine tends to quote. A brand sitting at fifty percent is either in the pool and losing the roll, which is a rewriting problem, or marginal to the pool, which is an authority problem. One blended number cannot tell you which, and prescribing off it is how a provider sells the wrong quarter of work with a clear conscience.

We are not the only people who think so, and the corroboration is worth more than our own argument because it comes from a company with no reason to help us. GrowthX publishes its product documentation publicly, and the page on how pages are scored carries the headings Health and Quality stay separate on purpose, Multiple axes separate problems a single score merges, and Incomplete is not a bad score. That is a direct competitor arguing, in its own docs, against the composite it would be commercially easier to sell.

The practical version for a buyer: ask any provider to show you the two numbers separately. How often are you in the retrieved set at all, and how often are you selected when you are. If the product cannot produce those separately, it is a chart rather than a diagnostic, whatever the score is called.

The routing question, and where we lose

Buy the tool if you have people who will act on it. An in-house content or search team, a marketing lead who owns the channel, and an existing habit of shipping changes on evidence. In that situation a subscription plus your own judgement beats an agency retainer on price and on speed, and one of the platform providers publishes a from-$6,000 monthly price for exactly that arrangement.

Hire a service if the measurement is going to sit unread. Most owner-led businesses we speak to are in this position: the number is not the constraint, the doing is. In that case the useful engagement is the diagnosis plus the execution, and the tracking is an input rather than the product.

There is a third answer that providers rarely give, so we will. Buy nothing yet if AI answers do not name businesses in your category. That is a measurable condition and it takes a few hours of instrument time to settle. Our free assessment exists partly to produce that conclusion, and the honest close on a diagnostic is that if the channel is not live in your market we say so and you spend nothing.

Where we lose, plainly. We have no revenue-attributed win yet. We hold every client citation share we have measured until it re-runs at three passes or more, so we cannot show you the number a dashboard would show you tomorrow. We do not measure Claude at all, because there is no ground-truth path to the consumer product, and we do not measure Perplexity either. We publish no guarantee while two of the six providers we audited do. We do not have the procurement documentation an enterprise buyer expects. And we will not post through persona accounts to manufacture the appearance of being recommended, which one provider in the audited set sells as a feature. If any of that is disqualifying, it should be, and the fastest way to find out is a call that costs nothing.