I Would Not Manage the Company to AI Share of Voice

I would not manage the company to a number called AI share of voice. I would manage it to whether an answer uses our pages as evidence: citation share, co-citation with the sources the engine already trusts, and how long a new page waits before its first citation. A mention can rise because the model listed a name. A citation moves only when a URL is doing work inside the answer.

The comparison is Share of AI Voice vs Citation: The Earned Media Metric. That page is the deeper resource. This essay is how I would use it as a founder when a vendor dashboard arrives with one score and a request to renew.

The June 2026 measurements split the two events

Nick Lafferty published a formula and benchmark for AI visibility metrics on June 9, 2026. The set was 48,589 citations in a single topic window, and the citation-share benchmark he measured was 0.50% (Nick Lafferty). That is a small share of a real denominator. A mention in a "companies to watch" line does not enter it. If the engine lists a brand and never uses a URL from that brand, share of voice can move and citation share stays where it was.

Maayan Zohar Basteker at Similarweb formalized the same split on June 11, 2026. She tracked 179 beauty prompts on ChatGPT in a Sephora case. Only 19 of those prompts generated citations at all. Sephora appeared in 3 of the 19, about 16% citation share on that cited subset. Her analysis of about 600,000 citation events found ChatGPT and Google AI Mode citing from different source pools (Similarweb). A brand can look present across prompts that never issue a citation. On those prompts the mention is an acknowledgment that a name exists.

Dan Taylor argued in Search Engine Land on June 8, 2026 that AI share of voice extrapolates from a small prompt set into an unbounded prompt space, and that the resulting assessment misleads (Search Engine Land). I would tape that objection to the dashboard. In classic search the keyword set could at least be listed. In an answer engine the prompt universe is not a list. Any fixed panel is a sample. A vendor who will not publish the sample is selling a ratio with a hidden denominator.

The output is unstable even when the prompt is held still. A 2026 study by Fishkin and O'Donnell found the chance that two responses produce the same ordered brand list for the same prompt was under 1 in 1,000 across nearly 3,000 runs. Lafferty's June data points the same direction: 95% of product titles appeared in under 30% of runs of the same prompt. A Tuesday share-of-voice reading is one draw from that distribution. I would not green-light a campaign because Wednesday's draw was higher.

Citation share moves too. The event it counts is narrower: the engine treated a URL as a source during retrieval. That is a stickier event than a name dropped during generation, because the next answer in the same topic has a source graph to reuse. I would still want the window, the engines, and the prompt list attached to the percentage. A stable kind of event can be badly sampled.

What compounding looks like in the published numbers

Ronn Torossian wrote on May 31, 2026 that in most categories 3 to 5 brands capture more than 50% of AI citations, that single-digit citation share marks a newcomer, that 10% to 25% is strong positioning for an established brand, and that above 25% is category leadership (Everything PR). I would use those bands as a map, with his source named, and I would not treat them as a quota. The useful part is the shape. The answer is a short list. A brand at zero citation share is absent from the set the answer is built from. A mention inside a list of twenty alternatives does not repair that absence.

Lafferty's time-to-first-citation figures make the calendar concrete: a median of 6.81 days from publication to first citation, a 90th percentile of 37.10 days, across roughly 900 pages first published between March and May 2026. A page that is built so an engine can reuse it can earn a citation inside a week. A year of those citations is a domain the retriever has already used. Share of voice has no equivalent feedback, because a mention does not enter the retrieval graph as a source.

After May 7, 2026, inline brand hyperlinks in AI answers moved from roughly 4% of responses to 22%, and those links drove more than 8 million measurable referral visits in Lafferty's account. That is a citation with a URL. A measurement system that increments the same counter for a disclaimer and for that hyperlink cannot tell me whether the May 7 change touched us.

Co-citation is the third number I would keep. Across more than 700,000 ChatGPT conversations, Lafferty found Edmunds and KBB co-appearing in 32% of answers that cite either one. Engines learn which sources reinforce each other. If our pages are never in that cluster, a rising mention rate is a name floating outside the evidence the answer is built from.

The scores being sold are presence scores

The DerivateX 2026 benchmark surveyed 50 B2B SaaS firms across 1,400 buyer-intent prompts and found an average AI Presence Score of 56.9 out of 100, with 44% of firms under 50. In concentrated categories the same write-up puts typical leader share of voice at 35% to 50%. In fragmented markets, 15% and above is described as strong (Cassie Clark). I would read those as presence. A firm at 35% share of voice can sit at 0% citation share if the engine says the name and never uses the firm's pages. The reverse is the firm I would rather be: fewer mentions, a larger share of the URLs the answer depends on.

A separate count, from LLM Pulse, says 89% of brands now appear in AI citations somewhere and only 14% measure those citations (LLM Pulse). The vendors selling share-of-voice dashboards are selling into that measuring minority, and many of the dashboards still blend mention and source. I would rather be in the 14% with a citation denominator than in the 89% with an unmeasured appearance.

Similarweb's June finding is the practical cut. ChatGPT and Google AI Mode draw on different pools. One blended share-of-voice number hides the engine where the company is a source and the engine where it is a name. I would keep the series per engine and per locale. I would not let a rollup decide a quarterly plan.

The terms I am using here sit in the glossary. The operating order I actually run is the one on Machine Relations: authority and entity clarity before a measurement layer that can flatter a mention.

How I would read a renewal against the June numbers

I would take the vendor's export and rebuild four lines from it before I discussed price. The lines are citation share, the prompt count that actually cited anything, co-citation with the sources already in the answer, and days to first citation on pages we published inside the contract. If the export cannot produce those lines, the renewal is a purchase of exposure and I would price it that way, or not buy it.

Lafferty's June 9 benchmark is the scale I would demand. A citation-share figure with no denominator is not comparable to 0.50% of 48,589 citations. A time-to-first-citation claim with no page count is not comparable to a median of 6.81 days and a 90th percentile of 37.10 days on roughly 900 pages published from March through May 2026. I would also ask whether their panel can see the May 7 change, when inline brand hyperlinks moved from about 4% of responses to 22% and accounted for more than 8 million referral visits in his account. A tool that increments once for a name and once for a hyperlink will report that week as a smooth trend. It was not a smooth trend. It was a change in how often the answer carried a URL.

Similarweb's Sephora cut is the template for a disclosed sample. One hundred seventy-nine beauty prompts, 19 of them producing any citation, Sephora in 3 of those 19. I can disagree with the category and still respect the fraction. A vendor who will only show me "16%" and will not show me 3 of 19 is hiding whether we are a large share of a tiny citing set. The 600,000 citation events underneath that work are what justify the second conclusion I would not let them average away: ChatGPT and Google AI Mode are different pools. I would renew per engine. A blended score that rose because one engine started mentioning us, while the engine our buyers use never cites our URLs, is a raise I would decline.

The instability numbers are why I would refuse a week-over-week share-of-voice goal. Fishkin and O'Donnell put the chance of two identical ordered brand lists, same prompt, under 1 in 1,000 across nearly 3,000 runs. Lafferty's 95% of product titles appearing in under 30% of repeat runs is the same warning in a product catalog. I would set the review cadence to a month, on a frozen prompt list, and I would require a run count large enough that one lucky draw cannot move the number across Torossian's bands. Those bands, from his May 31, 2026 piece, are the map I would print: 3 to 5 brands taking more than half of citations in most categories, single digits as a newcomer, 10% to 25% as strong for an established brand, above 25% as leadership. I would not turn the bands into a quota. I would use them to stop a team from celebrating 35% share of voice, the DerivateX leader band for concentrated categories, while citation share sits at zero.

DerivateX's 56.9 average presence score, 50 firms, 1,400 prompts, 44% of firms under 50, belongs in an appendix. Cassie Clark's write-up of the fragmented-market threshold, 15% share of voice and above as strong, belongs there too. LLM Pulse's 89% of brands appearing somewhere and 14% measuring the appearance is the market context for why the dashboard exists. It is not a benchmark I would manage to. The 14% who measure can still be measuring the wrong event. Edmunds and KBB co-appearing in 32% of answers that cite either one, across more than 700,000 ChatGPT conversations, is the event I would rather be in. If a renewal cannot tell me our co-citation with the three domains already trusted in our category, I would rather keep Lafferty's public benchmark as the external reference and run a smaller panel ourselves than pay for a presence score I cannot audit.

Taylor's June 8 objection stays the gate. If the prompt list is small, undisclosed, and then extrapolated to "share of AI voice," I would write that sentence into the rejection. The denominator is the thing I am buying. The chart is the decoration. A renewal that cannot show 3 of 19, or 6.81 days, or a per-engine split, is a renewal I would end even when the logo on the dashboard is familiar.

What I would do

I would open the vendor view and ask three questions in writing.

Does this number rise when the engine mentions us inside a disclaimer or a long alternative list? Does it rise when the engine uses a URL of ours as a source? What is citation share, what is co-citation with the three sources already trusted in the category, and how many days does a new page wait for its first citation?

If the first two questions have the same answer, the metric is exposure. I would keep the export for curiosity and take it off the scorecard that moves budget. I would fund pages that can be cited, on the queries buyers ask before they know our name, and I would review citation share on a fixed prompt panel each month, engine by engine.

Scope: commercial prompts in the category we sell, on the engines buyers in that category actually open. A branded prompt that already contains our name measures recognition, and I would report it in a separate column. I would not average it into discovery.

What would change my mind: a published panel, with the prompt list and the run count disclosed, in which mention rate predicts the next period's citation share tightly enough that acting on the mention is faster than acting on the citation. If that relationship showed up inside one engine and one locale, I would use share of voice there as a leading indicator and I would still keep citation share as the number I am willing to defend. Until then, the 0.50% benchmark on 48,589 citations is the kind of denominator I want, and a presence score of 56.9 is not a substitute for it.

FAQ

What would I count as a citation and what would I count as a mention?

A citation is a qualifying source slot that resolves to a URL. Lafferty's June 9, 2026 benchmark put a citation-share baseline at 0.50% across 48,589 citations. A mention is the brand string inside an answer, including disclaimers, roundups, and asides. I would store them as separate series. I would not let a vendor add them into one "visibility" score and then trend it.

Why would citation share compound if the answers themselves are unstable?

Fishkin and O'Donnell put the chance of two identical ordered brand lists, on the same prompt, under 1 in 1,000 across nearly 3,000 runs, and 95% of product titles in Lafferty's data appeared in under 30% of repeat runs. The generated list jumps. The retrieval decision is the part that can recur: a URL used as a source is available for the next answer in that topic. The median wait for a first citation, on about 900 new pages, was 6.81 days, with a 90th percentile of 37.10 days.

How would I keep a vendor honest about the denominator?

I would require the prompt list, the engines, the locales, the run window, and the rule that decides a mention versus a linked citation. Taylor's June 8, 2026 objection is the test. If the vendor extrapolates a small prompt set to "AI share of voice" without those fields, I would not put the number in a board deck. Similarweb's split, 19 citing prompts out of 179, is what a disclosed denominator looks like.

What presence score would I ignore?

I would ignore a single presence score that blends engines. DerivateX reported 56.9 out of 100 across 50 firms and 1,400 prompts, and leader share-of-voice bands of 35% to 50% in concentrated categories. Those figures describe how often names appear in that panel. They do not tell me whether any of those firms' pages were the evidence. I would ask for citation share on the same prompts before I treated 56.9 as a position.