Research · July 2026
The Four Shelves
There is no such thing as "the AI shelf." Four engines stock four different shelves — and the same brand can be on every one of one engine’s answers and missing from half of another’s, in the same week.
Earlier in The Invisible Shelf series, The Engines Disagree made a claim from a single observation: two AI engines looked at the same brand in the same category on the same day, and one recommended it while the other did not know it existed. Last week, Recognized, Not Recommended used the same phenomenon as supporting evidence, with the brands anonymized. Readers asked two fair questions: how common is this, and who are the brands?
This piece answers both. It is not a new claim. It is the dataset behind the old one.
The premise most teams carry into AI visibility is that there is a shelf — one place where AI-assisted buyers either find you or don't. The measurement says otherwise. There are four shelves, one per engine, and they are stocked differently. Not slightly differently. Differently enough that knowing your position on one tells you very little about your position on the others. Single observations like the one in The Engines Disagree are not edge cases. They are the structure.
What follows is one observation window — the week of June 28, 2026 — from the Envoyra observatory: 170 buying-intent questions across five software categories, submitted identically to ChatGPT Search, Claude, Gemini, and Perplexity. Appearance rates are reported per 100 questions in each brand's category.
The brand on every shelf — unevenly
Start with the least dramatic case, because it sets the baseline.
HubSpot, in the SMB CRM category, appeared in 92 of every 100 ChatGPT Search responses, 88 on Claude, 84 on Gemini, and 76 on Perplexity. Zendesk, in customer support: between 76 and 92 of every 100 on all four engines. ActiveCampaign, in marketing automation: 67 to 79 across the four.
That is what four-shelf coverage looks like: present everywhere, inside a narrow band. It is the position every brand assumes it has. In this window, few brands actually had it.
The brand on one shelf and half of another
Pipedrive appeared in 100 of 100 Claude responses to SMB CRM buying questions — every single answer in the window. On Gemini, 52 of 100.
Same brand. Same week. Same 25 questions. One engine considered Pipedrive part of the answer to everything; another left it out of nearly half.
Klaviyo, marketing automation: 77 of every 100 on Claude, 36 on ChatGPT Search. A buyer asking Claude about marketing automation encounters Klaviyo as a fixture of the category. A buyer asking ChatGPT Search meets it in roughly one answer in three.
Salesforce, in the customer-support category: 48 of every 100 on Claude, 8 on ChatGPT Search. Fireflies, meeting assistants: 88 on Claude, 40 on ChatGPT Search and Gemini.
One engine considered Pipedrive part of the answer to everything. Another left it out of nearly half. Same brand, same week, same questions.
The single-shelf brands
Then there are brands that, in this window, effectively existed on one shelf.
Brevo appeared in 47 of every 100 Perplexity responses to marketing automation questions — and 4 of every 100 on ChatGPT Search. Granola, a newer meeting assistant, appeared in 32 of every 100 Perplexity responses and zero on ChatGPT Search and Gemini. Read AI: 28 of 100 on Perplexity, zero or near-zero everywhere else.
For a buyer who uses Perplexity, these brands are part of the category. For a buyer who uses ChatGPT, they do not exist. Neither buyer knows the other's shelf is different.
Selected positions, week of June 28
| Brand (category) | ChatGPT Search | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| HubSpot (SMB CRM) | 92 | 88 | 84 | 76 |
| Pipedrive (SMB CRM) | 76 | 100 | 52 | 72 |
| Salesforce (SMB CRM) | 52 | 72 | 76 | 36 |
| Klaviyo (marketing automation) | 36 | 77 | 44 | 53 |
| Brevo (marketing automation) | 4 | 19 | 6 | 47 |
| Zendesk (customer support) | 76 | 92 | 80 | 76 |
| Salesforce (customer support) | 8 | 48 | 40 | 28 |
| Fireflies (meeting assistants) | 40 | 88 | 40 | 72 |
| Fathom (meeting assistants) | 16 | 56 | 40 | 44 |
| Granola (meeting assistants) | 0 | 8 | 0 | 32 |
Appearances per 100 category buying questions, June 28, 2026 observation window.
What this does to a single number
Average Klaviyo's four positions and you get a comfortable middle figure that describes none of its four realities. Averaging across engines does not summarize the data — it deletes the finding. The same is true of checking one engine: a team that tested Pipedrive on Claude that week would have concluded its AI visibility was total. A team testing on Gemini would have opened an incident.
Both teams would have been describing one shelf and believing they had described the store.
We are deliberately not explaining these gaps in this piece. Each engine assembles answers from its own sources, on its own cycles, and the mechanism behind any single brand's skew is not measured here. What is measured is the shape — and the shape repeats. Across all four June observation windows, the same brands showed the same skews: Klaviyo's Claude rate stayed at roughly double its ChatGPT Search rate in every week measured, Brevo's Perplexity rate stayed several times its ChatGPT Search rate, and Granola did not appear in a single ChatGPT Search response in any of the four weeks. The divergence is stable enough to be a property of the brand's position, not noise in the measurement.
The question to take away
The question "are we visible in AI?" has no answer, because it assumes one shelf.
The questions that have answers: Which engines name us, at what rate, on which buying questions? Which engines do our buyers actually use? And is the gap between our best shelf and our worst shelf widening or closing?
A brand that can answer those three has an AI visibility position. A brand with a single score has an average.
Methodology notes. Figures come from the Envoyra observatory run of June 28, 2026: 170 buying-intent prompts across five software categories, submitted identically to four AI engines through official APIs; 679 of 680 expected responses received. Rates are appearances per 100 prompts within each brand's category (25 prompts per category; 70 for marketing automation). Week-over-week persistence claims compare the four weekly June runs (June 7, 14, 21, 28) on the engines with complete coverage in each window. This is reported as an observation window, not a permanent ranking — positions move over time, which is measured separately in this series. One measured category (AI visibility tools) is excluded from the examples because Envoyra competes in it. Underlying response data is retained and auditable.
This article is part of The Invisible Shelf, a series by Envoyra co-founder and CEO Golara Serio on AI recommendation systems, brand discovery, and the measurement infrastructure being built in real time.