Our own AI visibility, published

What this does and does not show

It shows how often four generative engines named us in answers to twelve buyer-intent questions, on two dates. It is a small sample, and AI answers vary between runs, which is exactly why the same questions are repeated rather than swapped.

It does not show a clean movement between the two dates. Google AI Overviews blocked our sampling on 11 August and gave no reading, so that run pools three engines where the July run pooled four, and the number of usable answers changed with it. The headline change is between two different denominators. Comparing only the engines read on both dates is the fair test, and on that basis the honest summary is “roughly flat”, not a fall or a rise.

Neither figure measures pipeline. Being named in an answer is a discovery signal; it says nothing about whether a buyer visited, enquired, or was worth talking to. For clients we report those layers separately, and we hold ourselves to the same split. See the methodology for the three layers.

It also does not show that our approach is working yet. Between these two runs we published three rounds of content, Google impressions rose more than fivefold, and our share of AI answers fell. The straightforward reading is that publishing on your own domain does not buy you a place in someone else's shortlist, which is the argument we make on how to actually show up in ChatGPT and now have our own evidence for.

The method is documented on our methodology page and written up in full so you can run it yourself. Every claim we make about this field is graded in the evidence register.

How to read a share-of-voice figure — anyone's

The number at the top of this page is a ratio: answers that named us, divided by usable sampled answers. That is a decision, not a law of nature, and four other vendors would compute four other things from the same transcripts. Before comparing any two share-of-voice figures, work out whether they agree on these four points, because ours generally do not agree with anyone else's.

• The denominator. Is an unreachable engine counted as a zero, or excluded? We exclude it — the August run shows no reading for Google AI Overviews rather than a nought, because a block is missing data and scoring it as absence would understate us in a way that happens to flatter the next run.
• Whether a mention is weighted. Being named first in a shortlist of three is not the same event as being named ninth in a list of ten, and some vendors weight position into the headline figure. Ours does not: it counts named or not named, and position is reported separately so the two are not blended into one unarguable number.
• How many times each prompt was asked. A figure derived from one pass per prompt is a sample of one from a distribution. Ours repeats. A vendor unwilling to tell you the repetition count is telling you something.
• Who wrote the prompts. Prompts containing your own brand name will find your own brand. The twelve here are category questions a buyer would ask before knowing we exist, and they were frozen at the first run.

This is why we treat “AI Share of Voice” as a vendor coinage rather than a standard, including when the vendor is us. There is no standards body, no agreed formula and no cross-vendor comparability, which is graded contested on the evidence register. Compare a figure to the same figure measured the same way three months earlier. Do not compare it to someone else's.

What we would need to see before claiming this works

Two runs is a line between two points, and the line currently goes the wrong way. We published the equivalent table for the search side in page depth and position, where two clicks in three months is the headline figure and we left it in. What would change our mind here is not a bigger number in isolation — it is a movement we can attribute. Concretely: the same twelve prompts, at least three more runs, a rise that survives an engine being unreachable on one of the dates, and a plausible mechanism in the citation data rather than only in the headline ratio. If the share climbs while the cited URLs stay identical, something changed in the models rather than in our market position, and that is worth knowing but is not a result we can sell.

Publishing the interim state is the cost of holding that standard. We would rather show a flat-to-falling line with the method attached than a rising one with the questions quietly rewritten, which is the more common failure in this category and is invisible from outside. If you want the underlying transcripts for either run, the run identifiers are printed above — ask and we will send them.