AI visibility audit: what it checks, what it costs, and what it hides

· by boneLau

Short answer: An AI visibility audit runs a fixed set of buyer questions through several AI engines and records, for each answer, whether your brand was mentioned, whether the answer cited a source, and whether what it said was true. Most audits cost between $0 and $3,000; agency versions go past $10,000. The number of prompts is not what separates a good audit from an expensive one. The denominator is.

I'm writing this from inside the category. We sell a $49 one-off audit and a $39/month monitor, so read the pricing section with that in mind. Everything factual below comes from our own runs in October 2026 or from vendor pages we opened; where we don't know something, I say so.


What an audit actually records

An audit is not a score. It's a log. Per question, per engine, per run:

FieldExample
Question"Best AI visibility tool for a small SaaS?"
Engine + dateChatGPT, 2026-10-20
Mentioned?no
If mentioned: position#3
Cited our site?no — cited g2.com, reddit.com
Accurate?n/a (not mentioned)

Everything else — the score, the chart, the PDF — is a calculation on top of that log. If a vendor won't show you the log, you're buying a calculation you can't check.

The number that makes two audits incomparable

Every audit reports a mention rate. Almost none of them say what they divided by. There are three common denominators, and they produce three different numbers from the same week of data:

  1. All answers. 10 mentions out of 40 answers = 25%.
  2. Answers that name at least one brand. If 14 of those 40 answers recommended no product at all, the denominator becomes 26 and the rate becomes 38%.
  3. Answers where the prompt already contains your brand name. This measures something else entirely: when told you exist, does the model describe and recommend you? It does not measure whether a buyer would ever find you.

The third one is the dangerous one, because it's easy to produce and it looks great. We know because we did it. Our own dashboard showed 87/100 for brandscope.dev while a plain search for our category returned an AI Overview that did not mention us at all. The tool wasn't lying about its arithmetic. It was answering a question nobody had asked.

When we removed the brand name from the discovery prompts, our own mention rate went from 94% to 0/18 on those questions. Same product, same week, same engines. The full write-up is here.

So when you compare two audits, compare their denominators first. A 60% on denominator 3 and a 6% on denominator 1 can describe the same week.

What it costs, and what the price buys

Rough market as of October 2026, from vendor pages we actually opened:

SegmentTypical priceWhat you're paying for
DIY$0 + an afternoonYour own spreadsheet, 5 questions, 4 engines
Tool-generated report$0–$500Automatically collected data, light interpretation
Analyst-led audit$1,000–$3,000Someone reads the answers and writes recommendations
Agency / enterprise$5,000–$25,000+Wide competitor set, multiple markets, implementation help

A rule of thumb that held up across the pages we read: below roughly $1,500 you are buying tool output, not analyst time. That isn't automatically bad — a clean log is the part that is hard to fake. It's bad when the deliverable is dressed up as analysis.

Our $49 audit sits in the tool-output band. It produces the log, the denominator, the citation list and the accuracy check for one brand, once. It does not include an analyst reading your market for a week, and we don't claim it does. If that's what you need, buy the $1,000–$3,000 version from someone who does it properly.

One pricing pattern worth knowing: many agencies credit the audit fee toward the first retainer month. That makes the audit look free. It isn't. It makes the retainer the actual purchase decision.

Five things to check before you pay

  1. What's the denominator? If the answer is "all AI answers", ask how they handle answers that recommend nobody. If those aren't excluded, a category where the model answers without naming any product will look like your failure.
  2. Are the discovery prompts branded or unbranded? A branded prompt measures description accuracy. An unbranded one measures whether you exist. You need both, reported separately, not averaged.
  3. How many runs per question? One run is an anecdote — the same prompt can produce a different brand list the same day. SparkToro found the overlap between two runs of the same question was under 1%. Anything under three runs has a wide error bar, and a report that never mentions error bars is selling noise as trend.
  4. Which engines, which model versions? "ChatGPT" is not a stable target. Model updates change answers. An audit that doesn't date its run and name its engines can't be reproduced or compared six weeks later.
  5. Mentions or citations? Being named in an answer and being the source the answer cites are different outcomes, and they move at different speeds. Most reports blend them into one "visibility" figure. Ask for both separately.

What an audit cannot tell you

An audit is a snapshot. It tells you where you stand this week. It cannot tell you which fix worked, because:

  • AI answers vary enough that a single before/after pair proves nothing. A ±10 percentage-point swing between runs is normal at small sample sizes.
  • Research on answer stability is unkind. A St. Gallen study needed 33–94 repeats per prompt before a brand ranking stabilized; Peec estimates three months to a year to confirm a real change.
  • Citations rotate. One analysis found 40–60% of AI citations change month to month, with no action from you.

That's why we print a noise band next to every percentage, and why we tell customers to compare four weeks, not two. It's also why we don't call our action list proven. Nobody has measured whether those actions raise AI visibility — including us. The fixes we suggest are the industry's best guesses, and they're labelled that way inside the product.

Run a small one yourself first

Before spending anything, run five buyer questions through four engines and write down the answers. Twenty minutes, five columns, one rule: no brand name in the discovery questions.

The exact setup is here. If the result surprises you — most people find the unbranded number is far lower than the branded one — you'll know what to ask a vendor before you pay them. If you'd rather not do it manually, that's what the $49 audit is for.

Methodology

  • Engines: ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews — the five core surfaces we score. Earlier runs quoted in our benchmark post also included Qwen, GLM and Kimi; those are no longer in the product.
  • Sample: our own brand, brandscope.dev, weekly runs in October 2026. The run quoted above has 52 answers, 27 of which named no product at all.
  • Prices: vendor pricing pages read on 2026-10-24. Most are vendors' own marketing, not a neutral survey.
  • We sell audits and monitoring. There is no way to make that conflict disappear, so it's stated here instead.

Want your own baseline? Run one brand through the live question set — about 30 seconds, no account needed. Check your brand →

← All posts