Measure brand visibility in AI answers as a fraction and label both halves. Run a fixed set of buyer questions on one engine, on one date, from one place, then divide the answers that named the brand by the answers you ran: 12 of 40, or 30%. The 30% means nothing on its own. The 40, the engine, the prompt set, the date and the geography make it comparable.
The numbers in that example are illustrative, and so is every Acorn figure below; Acorn is a fictional brand. Rouse has not published client results. What we can show is how a rate is built, which is the part most visibility charts leave out.
What are brand mention rate, citation rate and AI share of voice?
They are three fractions with two different denominators, and the field guide's measurement section keeps them apart for that reason.
Brand mention rate is the share of sampled answers that name the brand. Citation rate is the share of sampled answers that link to the brand's own domain. Both divide by the same thing, the answers you ran, so they can be read side by side. An answer can recommend Acorn without linking to Acorn's site, and it can cite a page on Acorn's site while recommending something else.
AI share of voice divides by something different. It takes the brand's appearances and divides by every tracked brand's appearances in the same answers. Semrush's guide to measuring AI share of voice, published July 17, 2026, gives the plain version of the formula: your AI mentions divided by total AI mentions across all brands in your category, times 100. Its own Brand Performance report then weights that by how high the brand appears in each answer.
Ahrefs goes a different way. Its AI visibility metrics documentation defines AI share of voice as a brand's percentage share of impressions compared with other tracked brands, where impressions are Google search volumes summed over the prompts in which the brand appeared.
That is three vendors, three formulas and one label. A share of voice from one tool cannot be compared with one from another, and none can be compared with a mention rate. When we measure, the three visibility measurements keep their own denominators, which is why our note on how to measure Reddit and AI search visibility says they are never blended into one number.
Why does the rate move when nothing changed for the brand?
Because the denominator is the fragile half. Go back to the cover: 12 answers named the brand out of 40 sampled on one engine on one date, so 30%. Sample 10 answers instead and a single answer moves the rate by 10 points; at 40, one answer moves it by 2.5 points. A small sample makes the rate wide rather than wrong, and a report should say how wide.
Share of voice has a second way to move. Suppose Acorn appears in 12 answers, and its two tracked competitors appear in 18 and 10. That is 40 brand appearances, and Acorn holds 30% of them.
Add a third competitor that appears 20 times and the denominator becomes 60, so Acorn drops to 20% while its own count of 12 has not changed. Compare a share of voice month to month only when the tracked set is identical, and read the denominator even then.
The prompt set is the third lever. A set of comparison questions will name brands in most answers; a set of how-to questions will name brands in few. Change the mix and the mention rate changes with it, on the same engine, on the same day.
Vendors handle this differently, and the choice is rarely on the chart. Profound's product documentation says its visibility score divides the responses that include your brand by the responses that include at least one brand, so answers naming no brand at all fall out of the denominator. That is a different fraction from a mention rate over every answer run. We keep a fixed set of buyer questions per category, run the same set every time, and print a coverage line that says how many of them completed.
What do the public studies actually divide by?
Each published AI visibility study picked a denominator, and the choice explains most of the difference between their numbers. The table lists the studies we use, with each figure as its source states it.
| Study | Sample and window | Denominator | What it states about Reddit |
|---|---|---|---|
| Profound, citation patterns, June 5, 2025, updated August 2025 | 680 million citations, August 2024 to June 2025, on ChatGPT, Google AI Overviews and Perplexity | All citations in the corpus, per engine | 1.8% of ChatGPT citations, 2.2% of Google AI Overviews citations, 6.6% of Perplexity citations |
| Semrush, most-cited domains, November 10, 2025 | 230,000 prompts in weekly snapshots, July 14 to October 12, 2025, on ChatGPT search, Google AI Mode and Perplexity | Prompt responses, per engine, per week | Close to 60% of ChatGPT prompt responses in early August 2025, around 10% by mid-September |
| Ahrefs, most-cited domains, June 11, 2025 | About 76.7 million AI Overviews, 957,000 ChatGPT prompts and 953,500 Perplexity prompts, June 2025 | Mention share (how often the model chose the domain); reach weighted separately by Google search volume | 7.4% mention share in Google AI Overviews, third place; outside the top 10 on ChatGPT and Perplexity |
| Pew Research Center, July 22, 2025 | 900 U.S. adults, March 2025 browsing, results collected April 7 to 17, 2025, Google only | Source URLs listed in AI summaries, up to three per summary | Wikipedia, YouTube and Reddit together, 15% of listed sources |
Put the Profound and Ahrefs rows side by side. Both looked at Google AI Overviews in overlapping months, one reports 2.2% and the other 7.4%, and neither is wrong. Profound counts Reddit's share of all citations, and Ahrefs counts how often the model mentioned the domain. Our reference piece on how often AI answers cite Reddit goes through each of these studies in full.
The samples differ as much as the denominators. Profound's post does not say where its prompts came from; its product documentation says prompts are the queries Profound sends to answer engines daily, so the mix follows its customer base. Semrush's study does not state a language or country.
Ahrefs' Brand Radar methodology, updated February 26, 2026, builds its questions from Google's People Also Ask corpus and its keyword database, runs about 282 million AI Overviews queries and 14.5 million ChatGPT queries a month on a 90-day reporting window, and says that long-tail or niche prompts may not always be included. None of these samples is your prompt set.
What is a brand mention rate in ChatGPT worth without a date?
Less than it looks, because the rate for the same site on the same engine can fall by most of its value within weeks. Semrush's three-month study took weekly snapshots of 230,000 prompts from July 14 to October 12, 2025 and found that ChatGPT cited Reddit in close to 60% of prompt responses in early August before collapsing to around 10% by mid-September. The change was not uniform across engines; Reddit's frequency was more stable on AI Mode and Perplexity in the same weeks.
A site that quoted its August figure in October would have been off by roughly a factor of six without changing a word on its pages. A brand mention rate in ChatGPT measured in one week is evidence about that week. Semrush's share-of-voice guide adds that the same brand can have very different visibility across AI systems because each platform builds its answers differently, so the engine is a label too, and a rate averaged across engines has lost the information that made it useful.
This is what a dated sample means in practice: each measurement carries its observation date, and the next round runs the same questions so the two dates can be compared. Our note on what a Reddit marketing report should include covers where the dates belong.
A rate with no denominator is an opinion with a percent sign.
Why unmeasured is not zero
A zero says the engine was sampled and the brand did not appear. Unmeasured says no sample was taken. Those are different findings, and a report that writes 0 for both has thrown away the difference.
Google's own reporting shows how easily that happens. Search Console's generative AI performance report shows a tilde or a dash where a value is not available or not a number, and its help page states that those values will be zeros in the downloaded data. The same page notes that a site may see no data because it has not received enough impressions in generative AI features. An export therefore shows zero for a site that was never measured and for a site that was measured and found nothing, so keep the on-screen status next to any number you carry forward.
The report itself is worth having. Google announced it on June 3, 2026 as a dedicated view of impressions within generative AI features such as AI Overviews and AI Mode, broken out by page, country, device and date, and a note on the post says the insights reached all websites worldwide by August 31, 2026. It says nothing about ChatGPT, Perplexity or any other engine, so a brand with a strong Google report and no ChatGPT sample has one measured engine and one unmeasured engine, and should write it that way.
Rouse reports carry a status on every score: measured, partial or unmeasured. Referral traffic needs analytics access, so it stays unmeasured until that access exists rather than being recorded as nothing. A positive finding in one score stays positive when the score beside it is unmeasured.
What a citation or a mention does not prove
A mention proves that the brand's name appeared in an answer to a sampled prompt on a date. A citation proves that a link to the brand's domain appeared. Neither proves a visit, a reader, a preference or a sale.
Pew Research Center's July 22, 2025 short read analyzed the Google searches of 900 U.S. adults from March 2025, with results collected between April 7 and 17. In visits where an AI summary appeared, users clicked a traditional result 8% of the time, against 15% when no summary appeared, and clicked a link inside the summary itself in 1% of visits. A citation inside a summary was, in that sample, a link that almost nobody followed. Ahrefs says as much about its own metrics: its share of voice and estimated impressions model potential visibility rather than actual audience reach.
Causation is the other gap. A citation of a Reddit thread does not prove that a reply in that thread caused it; the engine may have retrieved the thread for reasons unconnected to the reply. Getting from a mention to an outcome takes the customer's own analytics, tagged links where a link is appropriate, and conversations with buyers who say where they found you. Our older piece on measuring Reddit engagement sets out that chain from published work to community response to business outcomes.
How to measure brand visibility in AI answers: a labeling checklist
Anyone asking how to measure brand visibility in AI answers can do it with a spreadsheet, and the discipline matters more than the tool. Before a visibility rate goes into a deck, check that it carries all of the following.
- The sample size, as a count. Write 12 of 40 and let the reader compute 30%. If the count is small, say so; the rate is wide.
- The engine, singular. One row per engine. Never average ChatGPT, Google AI Mode and Perplexity into one rate.
- The prompt set, fixed and kept. Keep the exact questions and run the same ones next time. A changed set is a new baseline.
- The date and the collection window. A rate is evidence about the day it was taken.
- The geography and language. Where the prompts were run from, and in what language.
- The denominator for share of voice. List the tracked brands. If the list changes, say so and draw no trend across the change.
- Unmeasured, written as a word. An engine or source you did not sample gets the word, never a zero.
- What the number does not show. Clicks, visits and enquiries live in analytics and in the engine's own reports, and belong beside the rate, never inside it.
The free Reddit visibility report applies this list to your own brand: three buyer questions run on a recorded date, every answer kept, and each score shown with its status and coverage line. It is a small sample by design, and it says so on the page.