July 24, 2026·Martin Endara

The Denominator Problem: Why Every AI Visibility Metric Ends in a Percentage Nobody Can Audit

AI visibility platforms report share of voice as a percentage. The denominator is proprietary and unauditable. Here’s why the industry stopped one step short of the number that actually matters.

In June 2026, Search Engine Land published something that most of the GEO industry would rather not have seen in print.

Dan Taylor, head of technical SEO at SALT.agency, laid out the case that traditional share of voice is effectively obsolete, and that many organizations replaced it with an equally flawed successor: AI share of voice. His argument is structural, not rhetorical. Unlike traditional search, where visibility could be measured against a known keyword set, the universe of possible AI prompts is effectively infinite. Which means the denominator in every AI visibility percentage you’ve ever been shown is a number the vendor invented.

Taylor is precise about how that invention works: vendors select a small, arbitrary subset of static prompts, run them through AI models behind the scenes, and aggregate those limited outputs into a representative global percentage, a process that measures share of voice within a contrived environment, presenting a closed sandbox as if it were the open web.

He’s right. And the fact that it appeared in Search Engine Land (a publication owned by Semrush) makes it more interesting, not less. The incumbent of the old measurement paradigm is publishing the autopsy of the new one.

But the article stops one step short. And that last step is the whole thing.

What the diagnosis gets right

The hidden denominator problem is real and it’s worse than most buyers understand. Taylor documents a case that should be required reading for anyone approving a GEO budget: when OpenAI shipped ChatGPT 5.0 in September 2025, the platform-wide volume of outbound citations dropped, and teams relying on LLM tracking dashboards saw a sudden sharp decline in their reported visibility, a decline that had nothing to do with brand relevance or marketing strategy. The model changed how it displayed sources. The dashboard called it a loss.

That is the failure mode of a metric with an unauditable denominator: it moves for reasons that have nothing to do with you, and you cannot inspect why.

Ahrefs supplies the second exhibit, and this one is cleaner because they ran the same study twice.

In their July 2025 analysis of 1.9 million citations from 1 million AI Overviews, 76.10% of AI Overview-cited pages ranked in the top 10. That number traveled. It went into decks. It underwrote the argument that AI citation was mostly a rankings problem wearing a new hat.

Then they ran it again at larger scale. Across the updated study, 37.9% of URLs cited in AI Overviews also appeared within the first 10 blocks, with the rest almost evenly split between positions 11–100 (31.2%) and beyond the top 100 (31.0%).

Seventy-six to thirty-eight. Same publisher, same metric, seven months.

And to Ahrefs’ credit, they don’t oversell the collapse. They attribute the gap partly to improved parsing methodology since the July 2025 study, which makes the two datasets not directly comparable, and partly to Google’s fan-out behavior. Read honestly, the number didn’t necessarily fall by half. The instrument got better at seeing.

Which is the point. When the instrument improving looks identical to the market collapsing, you do not have a measurement. You have a reading.

Where the industry stops

Taylor’s prescription is three replacement metrics: share of mentions, share of recommendations, and share of narrative. They’re smarter than what they replace, share of narrative in particular is a real insight, because a brand cited frequently but consistently described as a complex legacy system may have a high share of voice that is actively damaging its sales pipeline.

But look at the shape of all three. They are shares. Percentages. Each one requires a denominator, and each denominator is subject to the identical objection Taylor just spent 2,000 words making. He diagnosed the hidden denominator problem and prescribed three more hidden denominators.

This is not a criticism of one author. It’s the shape of the entire category. Every GEO platform on the market — the ones with hundreds of millions in venture funding and the ones with a landing page and a waitlist — reports in percentages. The competition is over which percentage. Nobody is questioning whether a percentage is the right unit at all.

A percentage is not a unit of decision

Here is the operational test, and it’s the one that matters if you’ve ever sat in the meeting where budget gets approved.

Your AI share of voice went from 12% to 18%. What do you do with that?

You cannot approve spend against it. You cannot compare it to the cost of the work that produced it. You cannot put it in a board deck next to a number that has a currency symbol. You cannot tell whether 18% is good, because you don’t know what the denominator was in either reading, and neither does the vendor’s own dashboard once the model updates.

A CFO does not evaluate a share. A CFO evaluates a number against a cost.

This is why the missing step isn’t a better percentage — it’s a different unit. The question a business actually needs answered is not what proportion of AI answers mention us. It’s how much revenue is exposed because they don’t.

That question has a currency symbol at the front of it, and a currency symbol is auditable in a way a proprietary denominator never will be. You can argue with the inputs. You can demand the traffic figure, the conversion rate, the average order value. You can substitute your own numbers and rerun it. Every assumption is on the table and every one of them is yours.

That’s the difference between a metric you can inspect and a metric you have to trust.

What this means operationally

Three things follow, and none of them require you to buy anything.

Treat every AI visibility percentage as directional. Taylor’s own framing is the right one: these metrics should be treated as directional signals rather than hard numbers. Use them to notice movement. Do not use them to justify spend, and do not let them into a board deck without that caveat attached.

Ask any vendor for their denominator before you ask for their price. How many prompts. Which ones. Chosen how. Refreshed when. If the answer is proprietary, you’re being sold a reading, not a measurement. That’s not necessarily worthless — but you should know which one you bought.

Convert to money yourself if nobody will do it for you. Traffic exposed to zero-click erosion, times your conversion rate, times your average order value. It’s arithmetic. It’s also the only version of this that survives contact with a CFO, and you can do the first pass in a spreadsheet before you talk to a single vendor.

Why we build it this way

Lotus quantifies revenue at risk in USD rather than reporting share of voice, and it does that for the reason above: a percentage with a proprietary denominator cannot be audited by the person paying for it.

The Bleed Model runs on traffic, an industry-varying bleed rate reflecting exposure to zero-click erosion, conversion rate, and average order value — with a deliberate conservatism factor of 0.44, because we report the minimum defensible scenario rather than the worst case. A number you can defend in a meeting is worth more than a number that sounds alarming and falls apart under questioning. The output is monthly and annual revenue at risk, labeled by source so you know which figures were measured and which were estimated.

And then Lotus generates the code — JSON-LD, llms.txt, structured data — because a diagnosis you cannot act on is just a more expensive version of the dashboard problem.

We don’t guarantee citations. Nobody credible does. We measure, we quantify in dollars, and we ship the artifacts.

Generative Engine Optimization (GEO) is the discipline of optimizing a website so that generative engines — ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews — retrieve, understand, and cite it when answering user questions. It does not replace SEO. SEO positions you in search engines; GEO determines whether models mention you at all. AI Revenue Protection is the category Lotus operates in: quantifying, in USD, the revenue exposed when generative engines answer a buyer’s question without citing you, and generating the executable code to close that gap.

We run it against your real domain, show you your bleed in USD, and hand you the artifacts ready to deploy.

Sources:

← Back to blog