Three Tools, Three Answers: The 7-Step AI Visibility Audit Every Marketer Can Run in 2026 Before Trusting a Dashboard

If you are a marketing director or analytics lead who has ever put the same brand into two AI visibility tools and gotten two numbers that could not both be true, this post is the test you run before you pick one. It takes an afternoon, it does not need a data scientist, and it tells you which of the numbers on your screen is decision grade.

TL;DR

• Three tools will give you three answers for the same brand. We have watched three published 2026 estimates of how much AI search cites YouTube land at roughly 1.8%, 23.3% and 23.1%. All three are about the same question.

• Most of the gap is method, not the engines. Different denominators, different engine mixes, single-shot sampling, and no range on anything. The engines are noisy too, but the tools are noisier.

• You can audit this yourself in seven steps. Pick one brand and one unbranded topic, run it in each tool, and ask each one the same seven questions. Most dashboards fail on step 2.

• A tool that cannot show its range is reporting noise as signal. That is the single test that separates measurement from monitoring, and it is visible on a demo floor in about thirty seconds.

What's broken in the category

More than 20 companies now sell AI visibility measurement, according to the IAB's August guidance, and each one reports a different number for the same brand with the same confidence. None of them is lying. They are measuring different things with different sample sizes and calling the results by the same name. Without a range on the number, a marketer has no way to tell which one to believe.

Same question, three published answers

Take a question every media planner cares about this year: how much of what AI search cites is YouTube?

Otterly AI published an answer on March 2. Across more than 100 million citation instances in 30 days, social platforms were 5.54% of all citations and YouTube was 31.8% of that slice. Work it through and YouTube is roughly 1.8% of everything cited. A few months later, 5W Public Relations published a second answer: YouTube at 23.3% of Google AI Overview citations, built from 46 million citations pooled from six other vendors' datasets. Then there is our own run: on the Google surface, youtube.com took 23.1% of citations, with a plausible true range of 19.2% to 27.2%. On the OpenAI surface in the same run, YouTube did not appear at all.

Roughly 1.8%. 23.3%. 23.1% on one engine and zero on another. Same question, same year.

Part of the gap is honest. Otterly's denominator was every citation on six engines; 5W's was Google AI Overviews alone. That is a disclosure problem, and the IAB's 4 P's framework exists partly to fix it. But two of the three published figures carried no range at all, and without the range you cannot tell a 23.3% that would survive a second run from a 23.3% that happened once.

Noise reported as signal

What the reports say: YouTube is 1.8% of AI citations. Or 23.3%. Each printed to a tenth of a point.

What the math supports: neither figure was published with a plausible true range, a repeat run, or a per-engine split. The 23.3% is a pooled number from six datasets with six different methods. The 1.8% is dominated by whichever engine cites the most, which on our data is Gemini at roughly 40 citations an answer against SearchGPT's six.

If you moved video budget on either number, you moved it on a reading you could not check. The IAB's guidance would call both of them exploratory, not decision grade.

One tool, three answers

It gets worse before it gets better, because a single tool can hand you three answers too, depending on how it rolls up engines.

In our own August run on the AI visibility tools category (123 buyer queries, 250 responses, 1,960 citations), g2.com was the second most cited domain overall at 11.4% of citations. Split it by engine and the 11.4% disappears. On Google's surface, G2 was cited once in the entire run: 0.1%. On OpenAI's surface it was 22.7%, with a range of 15.8% to 29.7%. The all-engine average is a number that is true on neither engine.

G2 at 0.1% on Google and 22.7% on ChatGPT. The dashboard that shows you 11.4% is not wrong. It is describing an engine that does not exist.

Most dashboards in this category pool citations across engines and report the pool. We compute every share per engine first and average the shares, so each engine counts once. That choice alone can move a brand's headline number by a factor of two, and almost no vendor discloses which way they do it. Ron Sielinski walks through the arithmetic in AI Visibility Metrics: How to Understand Them and How to Use Them, the first post in our measurement series, and the short version is that a pooled cross-engine share is a Gemini number wearing a cross-platform label.

Why the tools disagree, and why the engines do too

Think of flipping a fair coin 50 times. Sometimes you'll get 23 heads. Sometimes 27. Sometimes 30. Same coin. Just different runs. The variation isn't because the coin changed, it's because 50 flips is a small sample, and small samples bounce around a lot. AI visibility measurement at 50 prompts works the same way.

Now add the part the coin does not have. The engines themselves are random. In our first published study, we sent the same prompts to the same engines repeatedly and compared the domains they cited. The typical overlap between two runs of the same prompt was about 30% on Gemini, 40% on SearchGPT and 50% on Perplexity. The pages being cited were not changing; we fingerprinted them to check. The engines were.

So every visibility number is a sample statistic, and a sample statistic needs a range. That is the whole argument of AI Visibility Isn't a Number. It's a Range., and the one example from it worth carrying into a vendor meeting: on SearchGPT, for the topic of running gear, Tom's Guide took 9.5% of citations and Runner's World 6.0%. Put the ranges on and Tom's Guide spans 5.5% to 12.5%, Runner's World 4.0% to 8.0%. The whole Runner's World range sits inside the Tom's Guide range. Any dashboard would have shown a clear leader three and a half points ahead. The data shows a tie.

A tool that ran each prompt once, pooled the engines and skipped the range would have gotten a confident, precise, wrong answer. That is the default configuration of the category. More prompts alone won't fix it, and neither will faster dashboards, more engines, or overlaying agentic workflows on top of the same single-shot read.

The seven-step audit you can run yourself

You do not need our product to run this. You need one brand, one unbranded topic, trials of the tools you are considering, and the seven questions below. Where the 12-question scorecard we published in September grades the dashboard you already have, this audit is for the moment you are choosing between two or three.

Score each tool on the seven. When we last ran our broader 10-feature audit across five vendors in May, 46 of 50 cells came back No, three Partial and one Yes. Run the seven steps above and you will probably see a similar shape: most tools pass step 1, a few pass step 7, and almost none pass steps 2, 5 or 6. Those three are the ones that decide whether a number can carry a budget decision.

Marketer-Side Impact

Here is what the three-answers problem costs when nobody runs the audit. A tool reports 6% visibility on 40 prompts. A content sprint follows at roughly $750 per briefed piece, 60 pieces get produced, and the next read says 4%. Nobody can say whether the content failed or both numbers came from the same wide range. That is around $45,000 of production, a quarter of team time, and a figure you now have to explain in a QBR.

The audit costs an afternoon and whatever the trials cost. Every tool that fails step 2 is a tool whose numbers you would have been defending in a board meeting without a range to stand on. We estimate the category wastes about $2 billion a year this way, and the waste is concentrated in exactly the brands that bought on the headline number.

We ran it on ourselves, and the answer was zero

I should be upfront that we did not enjoy step 7. When we ran our own category through this process in August, IQRush appeared in zero of 250 answers and took zero of 1,960 citations. The zero was tight, with a range topping out at 1.5%, and the run had converged, which is what made it worth publishing. Profound and Peec AI own the category's answers right now. Four of the most-cited buyer guides are written by the vendors they rank first. We published the whole run, including the competitor table, on October 27: We measured ourselves at zero.

The point of including it here is the method. A measurement layer that cannot report an unflattering number about its own vendor is a marketing surface. We would rather show you the audit working against us than show you a case study.

Five questions to take into the demo

1. "Click that number. What range did the product compute around it, and where do I see it?"

2. "Is this a per-engine figure or a pool? Show me the same brand on Google alone and on ChatGPT alone."

3. "How many prompts, how many runs per prompt, and when did the ranking stop moving?"

4. "Show me a before-and-after for one brand. Where is the range on the difference?"

5. "Show me your own brand's numbers in your own category."

The fifth one is the tell. Every vendor in this space can run it. Almost none publish it.

Frequently asked questions

Why would three tools give different numbers for the same brand?

Three reasons, usually stacked. They sample differently (one run per prompt versus many), they aggregate differently (pooling engines versus averaging per engine), and they define the denominator differently (all citations versus citations on one engine). On top of that, the engines themselves return different sources on each run, with run-to-run overlap of 30% to 50% in our research. Any one of these can move a headline number by several points.

Which number should I trust if they disagree?

The one that comes with a range, a visible sample size, and a per-engine split. If none of them do, trust none of them for budget decisions and treat all three as exploratory, which is the IAB's own label for measurement under 50 queries or without reproducibility.

Do I need to run the full seven steps, or is there a shortcut?

Step 2 is the shortcut. Click the headline percentage and ask for the range. A tool that cannot produce one has not computed one, and every downstream feature (drift alerts, rank changes, week-over-week deltas) is built on a number with no known precision.

Isn't more prompts the fix?

Only up to a point. Ranges narrow with the square root of the sample, so cutting uncertainty in half takes four times the data. Past a certain size, more prompts buy very little. The better test is whether the tool can tell you when it has enough: whether the ranking has settled and the gaps between brands are bigger than the noise. We cover that in How to Tell If Your AI Search Visibility Number Is Real.

Can I run this audit on my current tool without a trial of anything else?

Yes, and you should. Steps 2 through 6 work on a single tool. The 12-question scorecard is the version built for exactly that. Use the seven steps when you are comparing, the scorecard when you are grading what you already pay for.

Todd Paris is co-founder and CEO of IQRush. Tracie Kambies is co-founder and COO. The video version of this audit, with one brand run through three dashboards on screen, is on the IQRush YouTube channel this week.

The falsifiable version: run the seven steps on your current tool. If it passes steps 2, 5 and 6, tell us which tool it is and we will say so publicly.

Book a walkthrough →

Published October 20, 2026. Last updated October 20, 2026.

Sources and verification

• Otterly AI, YouTube citation analysis, published 2026-03-02: social platforms 5.54% of 100M+ citations; YouTube 31.8% of the social slice.

• 5W Public Relations, 2026: YouTube 23.3% of Google AI Overview citations across 46M citations synthesized from six vendor datasets.

• IQRush own-brand run 41061b01, measurement date 2026-08-25: 123 queries, 250 responses, 1,960 citations, 525 pages, 248 domains; youtube.com 23.1% (19.2–27.2) on the Google surface; g2.com 0.1% Google / 22.7% (15.8–29.7) OpenAI / 11.4% all-engine mean.

• IAB, "Measuring Visibility in the AI Era," 2026-08-03: Exploratory / Directional / Decision-grade tiers; 20+ vendors; 4 P's.

• Sielinski, R. (2026). Quantifying Uncertainty in AI Visibility, arXiv:2603.08924 (run-to-run overlap; Tom's Guide / Runner's World example).

• Sielinski, R. (2026). From Stochastic to Stable, arXiv:2607.10341 (rank stability and sufficiency).

• Search Engine Journal, July 2026: AI Visibility Rankings Aren't Stable.

• IQRush 10-feature vendor audit, 2026-05-13: 50 cells, 46 No, 3 Partial, 1 Yes. "No" cells reflect strict absence from public documentation at the audit date; the comparison is re-checked weekly and cells move when vendors ship.