TL;DR: In July 2026 we asked ChatGPT, Perplexity, and Claude 30 buyer-intent questions about three real businesses in three different industries: a marketing agency (us), a B2B SaaS platform, and an eCommerce brand. That is 90 answers. The number of times any of the three was recommended in an answer that did not already contain its name: zero. AI assistants are recommending somebody in every category, every day. Right now it is almost never you, and the sources they build those answers from are not the sources most businesses invest in.
We run these probes for a living, which means we generate the raw data most articles about "AI visibility" only speculate about. This report publishes a quarter of that exhaust: what the engines answered, what they cited, and what actually moved. We count. We do not judge. Every number below has a date and a method attached, and we publish our own bad numbers with everyone else's.
- What we measured
- Finding 1: zero unbranded recommendations in 90 answers
- Finding 2: what AI cites instead of you
- Finding 3: the reviews question is a trap door
- Finding 4: we are in the dataset too, and our numbers are public
- Why this happens
- What to do about it (the five moves the data supports)
- Methodology and limits
What we measured
Between July 7 and July 13, 2026 we ran structured baselines for three businesses:
- Breadchaser.ai (us): an AI SEO agency. 10 prompts a buyer would ask about our category, from "best AI SEO agency for health and wellness businesses" to "what do people on Reddit recommend for AI SEO."
- A B2B SaaS platform (feedback and performance software; name withheld, they are a live engagement): 10 prompts a software buyer would ask in its category.
- A premium eCommerce brand (outdoor saunas): 10 prompts a shopper would ask before a four-figure purchase.
Each prompt went to three engines: ChatGPT, Perplexity, and Claude. Every answer and every retrieved source was logged as raw JSON. We classify each mention as brand-prompted (the question already contained the brand's name) or unbranded (the engine chose to name the brand on its own). That distinction is the whole report. A brand-prompted mention means the AI can describe you. An unbranded mention means the AI recommends you. Only the second one produces customers.
Finding 1: zero unbranded recommendations in 90 answers
The scoreboards:
| Business | Share of voice | Unbranded mentions |
|---|---|---|
| Breadchaser.ai (agency) | 6 of 30 | 0 |
| B2B SaaS platform | 5 of 30 | 0 |
| eCommerce sauna brand | 6 of 30 | 0 |
Every single hit, all 17 of them, came from prompts that already contained the brand's name. When the question was open ("best X," "which tool should I use," "what do people recommend"), the engines named somebody else every time. All three businesses have real customers, real revenue, and real websites. None of them existed to the engines as an answer.
If you take one number from this report, take that one. This is not a story about bad businesses. It is a story about a new distribution channel with its own rules, and almost nobody playing by them yet.
Finding 2: what AI cites instead of you
The engines tell you exactly how to win, because they show you their sources. Across our probes the retrieved-source lists were dominated by the same few surfaces:
- Reddit was the most-retrieved domain in the agency category (32 retrievals in one 10-prompt run) and near the top for the sauna category (28).
- YouTube showed up 20 times in both the agency and sauna runs.
- LinkedIn appeared 23 times in the agency run.
- In the sauna category, the single most-retrieved source (33 retrievals) was the market leader's own comparison listicles. The brand that dominates every unbranded sauna answer wrote the "best of" pages the engines quote, put competitors in them by name, and now effectively grades its own category.
Read that last one again. The most effective AI visibility asset we found in 90 answers was not a backlink profile or a domain rating. It was honest-looking comparison content published by a participant in the market. The engines treated it as the category's reference text.
This matches the strongest third-party data available: Whitespark's 2026 Local Search Ranking Factors survey (47 experts, 187 factors) added an AI visibility column this year, and the number one factor of all 187 was "presence of business on expert-curated best-of and similar lists." Citations that a human chose to write beat almost everything a tool can automate. (Whitespark, 2026)
Finding 3: the reviews question is a trap door
One probe deserves its own section. We asked the engines about the B2B platform by name, in the form a real buyer uses: "[brand] reviews." One major engine answered by recommending a competitor. The brand had too thin a review-and-listing footprint for the engine to say anything useful, so the engine filled the silence with the category leader.
That is the practical cost of an empty entity layer. It is not that AI says nothing about you. It is that AI answers questions about you with your competitors.
Finding 4: we are in the dataset too, and our numbers are public
We hold ourselves to the same probes. Our local-market baseline earlier this quarter scored 2 of 10 on ChatGPT and 1 of 10 on Perplexity. When we mapped every citation in our own category audit, the engines cited 69 distinct domains 119 times, and our own site was 5 of those 119 citations, about 4 percent. We published those numbers the day we measured them, and we are publishing these. The next edition of this report will show the deltas, up or flat.
Why this happens
The short version, backed by the Whitespark 2026 data: the signals that decide AI answers are not the signals most businesses have spent a decade optimizing. Google Business Profile signals carry about 32 percent of local pack ranking weight but only about 12 percent of AI visibility weight. Citations nearly double in relative importance for AI. On-page content, review breadth across platforms, and mentions on third-party pages carry the load. A business can be genuinely well optimized for Google Maps and be invisible to ChatGPT, and in our sample of 90 answers, that is exactly what all three businesses were.
What to do about it (the five moves the data supports)
- Get on the pages the engines already cite. Pull the retrieved-source list for your own category prompts and earn placement on those exact pages. In every category we measured, that list starts with Reddit, YouTube, and category listicles.
- Build the entity layer before you need it. Directory and review-platform profiles (G2 or Clutch for software and agencies, the health and wellness directories for clinics and studios, Bing Places everywhere because ChatGPT leans on Bing data) are what engines fall back on when someone asks about you by name. Empty profiles hand the answer to a competitor.
- Publish honest comparison content. The sauna market leader proved the play. It only works if it is actually honest: named competitors, real criteria, your own gaps shown. Google has started penalizing self-serving fake listicles, so the transparency is not optional decoration, it is the mechanism.
- Get reviews on more than one platform. The engines cross-reference. In the 2026 data, reviews are the only signal category in the top two for both traditional local ranking and AI visibility.
- Measure share of voice, not rank. AI answers are non-deterministic. Track how often you are named across a fixed prompt set over time. That is the number this report is built on, and the number we re-publish every quarter.
Methodology and limits
Three businesses, one per vertical (services, B2B SaaS, eCommerce). Ten buyer-intent prompts per business, fixed before the runs. Three engines per prompt (ChatGPT, Perplexity, Claude; Gemini was excluded this quarter for API quota reasons). Runs dated 2026-07-07 through 2026-07-13, raw responses and retrieved-source lists logged at run time. A mention is counted once per answer. Brand-prompted and unbranded mentions are never summed together. This is a small sample published because almost no primary data exists at the category level; it is a census of our own probe exhaust, not a survey of the market. Reproduce it: pick 10 prompts, ask three engines, count.
The Q4 edition ships in October with the same prompt sets, so the deltas mean something.
Want the same probe run on your business? The free AI Visibility Score asks the engines about your market and shows you who they name instead of you, with receipts.
