What to expect from an AI search visibility audit
Last updated October 14, 2026. Written by the Third Coast AI team.
A good AI search visibility audit answers four questions with evidence: which companies ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity and Copilot name when your buyers ask, which pages and sources those engines are citing, which decision factors they use to choose, and what specifically your site and your third-party record are missing. It should take two to four weeks, it should be built on a prompt set written around your buyers rather than a keyword list, and it should end in a ranked plan, not a score. This page describes what belongs in one, what the weak ones leave out, and the questions to put to any vendor, including us.
We wrote it because "AI visibility audit" now means anything from a free dashboard screenshot to a month of real work, and the price range reflects that. If you are about to pay for one, the checklist below will tell you within ten minutes which kind you are being offered.
Which parts are essential?
| Component | What it should contain | Why it matters |
|---|---|---|
| Prompt bank | 60 to 120 prompts across visibility, decision-factor, competitor and fact-check types, written for named buyer personas | The prompts are the measurement; a keyword list dressed up as prompts measures the wrong thing |
| Multi-engine runs, repeated | ChatGPT with web search, Gemini with grounding, AI Overviews, AI Mode, Perplexity, Copilot; each prompt run more than once | Only about 11% of cited domains overlap between ChatGPT and Perplexity, and single snapshots are unreliable |
| hidden-search captures | The hidden sub-queries the engines issued behind each prompt, grouped into topics | Tells you what to write; most tools do not show it |
| Decision factors | The criteria the engines use for your category, in their words, with frequency | This is the RFP you are answering |
| Proof audit | Your pages checked against each factor: stated, proven, missing | Turns the factors into page-level work |
| Citation reads | The pages the engines cite for each topic and the structural patterns they share | The brief for your pages comes from here |
| Fact-check log | Everything the engines currently get wrong about you | Often the fastest wins, and the most embarrassing misses |
| Ranked plan | First 90 days of fixes, briefs and off-site work, ordered by expected effect | An audit that ends in a score has not finished |
Can you tell a weak audit from a strong one before paying?
Usually from the sample deliverable. A weak audit shows a visibility percentage, a sentiment gauge and a list of competitors, all generated by a tool in an afternoon, with recommendations like "add FAQ schema" and "publish more content." Ahrefs' difference-in-differences test on 1,885 pages found FAQ schema made no measurable difference to AI Mode or ChatGPT citations, so a recommendation like that is a sign the vendor is working from a 2024 blog post rather than from your data. A strong audit shows you the actual answers with citations, the sub-queries, a factor list you recognize from your own sales calls, and a page-by-page table of what your site fails to say.
The other tell is the prompt bank. Ask to see it. If the prompts are keywords with "best" in front of them, the audit will measure keyword visibility and call it AI visibility. If they read like things your customers actually say, including the awkward ones ("is it worth paying more for a local company or should I just use the chain"), you are in good hands.
What should the timeline look like?
Two to four weeks is normal for a single brand in a single market. Week one is intake and prompt-bank construction and needs about two hours of your time; week two is the runs, the citation reads and the proof audit; week three is the plan and the readout. Anything promised in 48 hours is a tool export, and anything that takes two months is probably bundling content work into the audit. Our own audit runs three weeks and adds a week per additional market or brand.
The audit we would want if we were the buyer shows us the engine's actual words, the questions it asked behind our back, and the sentence on our website that should have answered them and didn't. Everything else is a chart.
Third Coast AI
Which numbers should the audit be able to explain?
A competent vendor should be able to put your results in context without hand-waving. BrightEdge's tracking now puts AI Overviews on nearly half of Google queries, and Semrush measured a 71% rise on commercial searches between November 2025 and April 2026, so the audit should say where your category falls. Pew Research measured click-through at 8% with an AI summary against 15% without, so the audit should connect visibility to the click you are losing or keeping. Ahrefs measured Google top-ten overlap with AI Overview citations at 38% in March 2026, down from 76%, so the audit should not assume your rankings are a proxy. And Semrush's finding that AI-search visitors convert at 4.4 times the rate of classic organic should shape how the vendor talks about value: fewer visits, better ones.
Are there questions worth asking every vendor?
- Can I see the prompt bank before you run it, and will I own it afterward?
- Which engines do you run, with retrieval on or off, and how many times per prompt?
- Do you capture the engines' hidden searches, or only the final answers?
- How do you extract decision factors, and can I see them in the engine's own words?
- Will you audit my pages against those factors, page by page?
- Which pages do the engines cite for my topics, and did you read them?
- What does the engine currently get wrong about my company?
- What are the first five things you would change, in order, and why those?
- Which of your recommendations are supported by published research, and which are your judgment?
- If I continue with you, is the audit fee credited?
What happens after the readout?
You should leave with the plan, the prompt set, the raw answers and the freedom to do the work yourself. Many companies hand the briefs to an in-house writer and the entity list to whoever handles PR, and come back for a re-run in three months. Others continue into a program: on-site work through something like our GEO content program, off-site through a citation and entity program. Either path is fine. What should not happen is that the audit sits in a folder, because the engines keep moving: half of Perplexity's citations are to content under 13 weeks old, and a competitor who publishes the decision-factor page you did not write will take the slot.
How much does one cost?
Prices vary widely with the amount of human reading involved. Tool-generated reports are sometimes free or bundled with a subscription (OtterlyAI, for example, publishes plans from about $29 to $489 a month (checked October 2026)). Consultancy audits of the kind described here are fixed-fee projects; ours is quoted after a short scoping call and is credited against the first month of a program. The pricing guide goes into how agencies and tools price the work.
Frequently asked questions
Two to four weeks for one brand in one market. A 48-hour turnaround is a tool export; two months usually means content work is bundled in.
A dashboard reports how often you are named. An audit explains why, by capturing the hidden searches, the decision factors, the cited pages and the gaps on your own site, and it ends in a ranked plan.
It should cover ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity and Copilot at minimum, because their citation sources barely overlap. The weight given to each depends on your category and buyers.
You should. Ask before you buy. Ours is delivered with the raw answers so you can re-run it yourself.
A good one gives you briefs or at least the citation reads and decision factors a brief is built from. If the recommendations are generic ("add FAQ schema"), the audit did not go deep enough.
As a prompt to act, yes. As a plan, no. Free reports are tool exports with a score, and Bing Webmaster Tools now offers a free first-party AI Performance report that is worth checking regardless.