AI Visibility Audit Deliverables for a UK Agency: The Five Artefacts, and Why the Zero Column Is One of Them

An AI visibility audit deliverable is five artefacts, and two of them are the ones agencies leave out: the column that reads zero, and the log of findings the audit withdrew. Our own British cells are the demonstration. Across 420 United Kingdom readings on 16 October 2026, from 630 questions put to ChatGPT and Gemini, a page of ours was among the sources in 3 readings. All 3 were on ChatGPT. The Gemini column reads 0 of 210. Hand a client the first number without the second and the report cannot be re-run, which is the only test a client report has to pass.

We sell this measurement and the publishing work that follows it, so the uncomfortable items below are ours.

Artefact one: the question set, as a file, versioned and dated

The set is the denominator. Change the wording and the number moves for reasons unconnected to the client's visibility, so the file carries a version number and a date and the client keeps a copy. The client's brand name appears in no question and in no wrapper.

This is the cheapest item to satisfy and the one most often missing. 78 of the 102 build spec conversations we held on 15 October 2026 came back asking for the prompt set to be made public so it can be run again. A set a stranger can run is the difference between a report and a claim.

Artefact two: the row level log, with attempted and usable both printed

One row per reading, and two counts at the top of it. On 15 October 2026 we set out to take 612 readings and kept 597 of them. On 16 October the figure set out for was 1,260 and the figure kept was also 1,260. The second is the stronger claim because the target total sits beside the kept total, and a report showing only one of the pair has removed the part a careful client asks about first.

The log also states the unit. One citation here is one domain inside one reading: four appearances of the same domain in a single answer are still one. Score the same rows by raw URL occurrence instead and the total climbs to 36,576. Neither is wrong, the two are not comparable, and the unit belongs on the first page.

Artefact three: the table, with one row per engine and the zeros left in

Average the two engines together and the resulting figure belongs to neither of them. Our six cells from 16 October 2026, each over a denominator of 210, read 1 on United States ChatGPT, 3 on British ChatGPT and 2 on Indian ChatGPT, against 0 on Gemini in all three markets. The 15 October 2026 run added 291 more Gemini readings with the same outcome, putting that column at 0 out of 921.

The zero is not a blank and not a failure of the work. The two surfaces concentrate their sourcing very differently, which the per engine domain tables in our benchmark set out in full, so one of them has a short entrenched list to displace and the other has none. A British client reading a zero in the Gemini column is reading the starting position of that surface. Which metrics survive into a client scorecard and which mislead transfers directly from search reporting, and the SEO metrics worth reporting and the ones that mislead is the nearest reference for it.

The table also needs genuinely British rows in it. gov.uk carried 57 of the UK readings on 15 October 2026 and ico.org.uk 28, and neither shows up in the American set at all. Small British agency domains took dozens of readings apiece alongside them, and a UK report with no .co.uk in it is describing the wrong field.

Artefact four: the withdrawal log, which is the hardest one to hand over

68 of those 102 build specs asked for failures and negative results to be shown, and almost nobody in this category publishes them. Here is ours, from 220 dated fetches on 16 October 2026.

Two hosts looked like they were refusing named AI crawlers with an HTTP 403 while serving a desktop browser. A same day re-run in reversed agent order agreed on every status, which looked like proof and was not: hours later, from a different connection, both returned HTTP 200 and the whole article to almost every agent, browser included, and one of them answered HTTP 429. They were rate limiters wearing a 403.

So two of four findings did not survive, and the rule they produced is the useful part: a 403 under a bot agent counts only when it repeats in a separate session hours later, and a 429 anywhere in a set puts the whole set in doubt. Two did survive, on both readings: help.openai.com/en/ refused GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot, and refused a request with no agent string as well, and theverge.com refused PerplexityBot and ClaudeBot. A client who has seen you remove your own row will believe the rows you keep.

Artefact five: the correction log and the next reading date

Every one of the 102 build specs wanted three things on a page: when it was written, when it was last checked, and how often the checking happens. The correction log is what makes those honest, because a dated page that quietly changes is worse than an undated one. Two entries from ours.

We published a vendor own site share of 91.3% and it was wrong, because the sorter behind it misfiled platform documentation as marketing. The corrected figure is 85.6%, every source now passes through a tested module that will not start unless all 24 of its named examples land correctly, and the old figure stays named on the page rather than deleted.

The second is about our own website. Fetched as GPTBot, our pricing page served rupee figures and no pound figure at all, because the currency conversion ran in the browser and a crawler sends none of the signals the detector looks for. It was fixed the same day and now carries 5 pound figures, 5 euro, 15 dollar and 31 rupee in the visible text on every request. Our own crawler check could not see it: run from our own network both the browser and the crawler received rupees, so there was no difference to find. If a British client is quoted in pounds anywhere, that is the check to run on their pricing page first.

Three absences belong in the same log. No rate by trade exists in either run, because the questions were written without one. No British city appears as its own figure in the 16 October run, since those readings belong to a separate ten city scan of 120. And only two of the five surfaces the product reads were measured here, ChatGPT and Gemini, with Perplexity, Claude and Google AI Overviews absent from every figure in the pack.

Questions a British agency gets about an AI visibility audit deliverable

How do I present a round where nothing moved?

Identically to a good one, numerator showing, so the client can count the readings that moved. On a denominator of 210, a single reading is under half a percentage point. Anything finer than one reading is movement the data cannot contain.

Do I have to show the client the findings I withdrew?

Yes, and it is the strongest page in the document. Two of our four access findings on 16 October 2026 turned out to be rate limiters rather than refusals, and publishing that produced a rule worth more than the four findings were. An audit that has never removed one of its own rows has not been checked.

Should the tool cost appear in the audit deliverable?

As a line with its VAT basis beside it, yes. Software in this category is published in dollars, so a pound figure is somebody's conversion, and whether UK VAT is charged by the supplier or accounted for by the client depends on where the supplier is established. Have the supplier state the basis in writing and keep it in the same row as the amount. A client who is not registered pays the gross figure, so that is the number their comparison needs.

How often should the audit be re-run?

On a fixed interval set before the first result is seen, with the question set unchanged. Our two passes agreed with each other. The earlier one put us in 3 of 597 readings and the later one in 6 of 1,260, while the self published share moved from 85.6% to 82.0% and the directory share from 0.8% to 0.7%. Agreement between two readings is what a client is paying for.

The deliverable a British client can argue with

Ask an assistant today about a British supplier and you get a fluent answer built from pages that print no sample size and name no engine. That is the field this deliverable competes in, which is why the five artefacts are weighted towards what a stranger can check rather than towards presentation. We are the vendor here. AI Knows Us charges for the reading on five surfaces and for the writing that follows, which is why our own British cell appears as 3 of 420 with a zero beside it. The 597 reading benchmark across ChatGPT and Gemini in three markets is the worked version of all five, including the parts we have since corrected.