An AI visibility audit that
measures instead of guessing.
Most audits in this category inspect your pages, compare them to a checklist, and infer what AI probably thinks. This one puts an instrument on each of the four things that can actually be measured, and tells you which of them is your bottleneck.
Four questions, four instruments.
Can AI agents actually read your site?
Server response and rendered word count under each major agent — Googlebot, GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot. The failure I find most often is a site that returns a full page to a browser and a near-empty shell to a bot, which makes every downstream question moot.
You get: A pass/fail table per agent, with the word count each one receives.
What are assistants already fetching?
A logger goes on your server, ahead of the page cache, and records every AI request per URL — separated into background crawling, training collection, search indexing, and fetches made while composing an answer for a real person. That last category is the one no third-party tool can see, because it only exists in your log.
You get: Your live-fetch list: which URLs assistants reach for, how often. On one client site that was 207 URLs out of 1,703 crawled.
Do you have a citable surface at all?
Assistants quote pages that make a factual claim with a number attached. Service and location pages almost never qualify, which is why a site built entirely of them earns no citations no matter how well written it is. I inventory what you have that is quotable, and what proprietary data you are sitting on that nobody in your industry has published.
You get: A ranked list of pages to restructure, and two or three data assets your own records would support.
Would you know if AI sent you a customer?
Analytics undercounts this channel by roughly seventy times, because the answer — and in home services the phone number — is delivered inside the chat. I check what your analytics is capable of attributing, where first-touch is being lost, and what it would take to carry an origin into your CRM next to the invoice.
You get: The specific gaps in your attribution chain, in the order they should be closed.
Why four rather than one: these numbers move independently. On one client site, presence in AI answers grew roughly forty-fold in a window where domain authority briefly fell and organic traffic stayed flat. A single-number audit would have missed which lever was working. The four numbers explained →
What lands in your inbox.
- A written report, in plain language, with every claim traced to the measurement that produced it.
- The four tables above, as data you keep — not screenshots of a dashboard you lose access to.
- A prioritised list of fixes, ordered by result per hour of work, with the ones I think are not worth doing marked as such.
- The logger left installed on your infrastructure, with the log in your account. You keep it whether or not we work together afterwards.
- A 30-minute call to walk through it, if you want one. The report stands on its own if you do not.
What is not included, said up front
- Implementation. This is a diagnosis; the fixes are a separate, priced engagement — and some of them you can hand to your existing team.
- A prompt-sampling subscription. If you want share-of-voice tracking, I will tell you which tool covers your engines and roughly what it costs, and you buy it directly rather than through me at a markup.
- A guarantee of citations. Nobody can sell that honestly. What I can tell you is whether you currently have anything citable, which is the part most sites fail.
- A hundred-page deck. The last one ran to a handful of pages because the findings were the point, not the volume.
If the audit shows the work is worth doing, the ongoing arrangement is a separate decision with published prices and no obligation attached to the audit. Pricing →
Questions about the audit.
A measurement of four separate things: whether AI agents can read your site, what they are already fetching from it, whether you own anything they would quote, and whether you could tell if one of them sent you a customer. Those four are measured by four different instruments, they move independently, and most reports sold as AI visibility audits cover only the third one — usually by inspecting your pages and guessing.
Mine starts at $1,500, published on the pricing page, and includes the logger install so the second layer is measured rather than estimated. For context on the market: published ranges for one-off generative-engine-optimization audits run roughly $5,000 to $15,000 at agency and engagement grade. Verify that against whoever you are comparing me to — this market reprices constantly.
The site-side work is a few days. The honest constraint is the logger: it needs about a week of live traffic before the live-fetch list means anything, and I would rather hand you a real week than a same-day report built on nothing. So plan on two weeks from install to report.
Then that is the finding and you get it in writing, along with which of the four layers is the reason. That outcome is common and useful: a site with no citable pages does not need a monthly retainer, it needs two data pages. I would rather tell you that than sell you twelve months of content.
Analytics read access and a way to install the logger — on WordPress that is a plugin, elsewhere it depends on the stack. No write access to anything, no changes to your site during the audit. If you cannot grant server access, layers one, three and four still work and I will say plainly that layer two is estimated rather than measured.
No. Home services is where I test the method, because it is the hardest configuration — the conversion is a phone call, the job is created by hand with no source field, and nobody links to a repair company. A method that produces a traceable job there transfers to easier setups. The vertical is a stress test, not a restriction.
Find out which of the four layers is stopping you.
Send me the domain and I will tell you for free what I can see from the outside, before anyone pays anyone. That answer is often enough on its own.