Forty-one of 108 businesses received no website citation in a recent study. That is not evidence that 38% of businesses are invisible to AI. It is evidence that a narrow measurement can become a much bigger sales claim on its way to a pitch deck.
Bodhi & Co ran 422 buyer questions against 108 businesses across August and September 2026. Its report counts businesses whose own web address never appeared in an answer’s sources or text. A brand named throughout an answer without a link does not qualify as cited. Missing URLs and missing businesses are not the same thing.
The authors make that distinction themselves. Their follow-up piece separates how often a business is cited, named, and recommended. Those are different outcomes. Treating one as a substitute for all of them is the problem.
The denominator needs attention, too. Only 21 businesses were tested on all four engines; 73 were tested on Claude alone. A business uncited by Claude and a business uncited across multiple engines enter the same headline count, despite having different opportunities to appear.
The reported percentage describes this panel. It does not establish an AI-invisibility rate for businesses generally. I have spent enough time around paid-search dashboards to recognize the move: measure something narrow, give it a broader name, and let the label do the selling.
Different denominators answer different questions
The report’s other counts make the distinction easier to see. The business’s own website appeared as a source in 132 of 422 questions, while 13 of the 108 businesses were cited on every question they faced.
| Measure | Numerator | Denominator | Rounded rate |
|---|---|---|---|
| Businesses with no website citations | 41 | 108 businesses | 38% |
| Questions where the business’s site was cited | 132 | 422 questions | 31% |
| Businesses cited on every question they faced | 13 | 108 businesses | 12% |
A business-level rate and a question-level rate are not interchangeable. The first asks how many firms accumulated no citations across their tests. The second asks how often a question produced an own-site citation. The last asks how many firms had a perfect citation record across the questions put to them.
Exposure matters. A firm facing several difficult buyer questions has a different test from a firm facing one. “Cited every time” also means something different depending on how many times the business was tested. The denominator is not administrative detail. It tells me what the result means.
None of those measures establishes recommendation share. An engine can name or recommend a business while citing a review site, directory, or Reddit thread instead of that business’s website.
Slate’s benchmark of 4,500 answers across ChatGPT, AI Overviews, and Perplexity illustrates the separation. Among roughly 10,000 cited sources, 41 of the 50 AI companies studied received no citation to their own website. Yet the typical company still appeared in about one in ten relevant answers. That is another dataset, not a correction to Bodhi & Co’s rate. Its value here is showing why uncited cannot stand in for unnamed.
Nor does any of this establish revenue. A visibility test records what an engine emitted during a test window. It does not establish whether a buyer read the answer, trusted it, or bought.
Before I accept an invisibility rate, I want to know what its denominator counts: businesses, questions, answers, or sessions. Then I want to know whether an uncited-but-recommended business counts as invisible. If it does, the label is already failing.
Uneven engine coverage limits the headline
The cleanest comparison in the Bodhi & Co report is the group tested across every engine. Even there, the result depends on which engine is measured.
| Engine | Questions citing the business’s site | Questions tested | Rounded citation rate |
|---|---|---|---|
| Gemini | 19 | 109 | 17% |
| Claude | 18 | 110 | 16% |
| Perplexity | 17 | 109 | 16% |
| ChatGPT | 4 | 109 | 4% |
Within that group, 12 of the 21 businesses received no citation from any of the four engines. That is a useful observation about those businesses under those tests. It is not equivalent to receiving no citation from Claude alone, yet both observations contribute to the overall uncited count.
The panel can reveal patterns worth testing; it cannot supply a market-wide invisibility rate. These businesses came from one agency’s audits, not a probability sample of the market. Uneven engine coverage adds another constraint. The headline percentage compresses both limitations out of sight.
The report also states a limitation that deserves more prominence: it queried developer APIs with live web search enabled, not consumer apps. It explicitly warns that app answers can differ.
That distinction is not cosmetic. First Answer ran the same question set through the ChatGPT API and interface, with 1,000 queries on each surface. One tracked brand appeared in 11.1% of API responses and 92.6% of browser responses. That result does not give me a universal adjustment factor. It tells me that substituting one surface for another can change the observation substantially.
Other research points in the same direction, though it measures different things:
- Wang et al.’s EMNLP study found API evaluations scored 3.4 percentage points higher in accuracy and 2.1 points higher in test-retest agreement across seven systems. Those are evaluation differences, not brand-visibility rates.
- Surfer’s analysis of 13,779 answers found ChatGPT’s API named 13.8 brands per answer against 7.9 in the consumer app. API answers were also about 2.05 times as long.
I cannot use those studies to recalculate Bodhi & Co’s results. The prompts, samples, and measures differ. I can use them to reject the assumption that API output automatically represents the answer a buyer sees in an app.
Ask which surface produced the number before asking whether the number improved.
A visibility report needs separate, inspectable measures
The reporting fix starts with definitions, not a more impressive dashboard. CitedMetrics publishes a worked rubric that separates the outcomes and links findings to dated raw answers. Its method repeats every prompt three times per engine in clean sessions.
| Measure | What the report should identify |
|---|---|
| Mentioned | The brand appears in the answer. |
| Cited | The brand’s own website appears as a source. |
| Recommended | The engine presents the brand as a pick, not merely a passing mention. |
| Invisible | None of those outcomes occurs. |
Mentioned, cited, and recommended need separate reporting, not one blended score. An answer can name a business, cite its website, and recommend it. A report needs to show those distinctions rather than hide them inside a proprietary calculation.
Then comes the measurement label. I would put the following disclosures in the contract, not leave them for a methodology call after the invoice:
- Prompts: the exact wording and the topics they cover.
- Engines and surfaces: which engine, consumer interface or API.
- Repetitions: observation counts per prompt and per engine.
- Timing: the date window for the runs.
- Session state: clean or logged-in, fresh or carrying memory.
- Search behavior: forced search or the engine’s natural behavior.
- Evidence: dated raw answers supporting the reported result.
A survey of 163 practitioners calls the blended visibility score “the easiest thing to sell and the hardest thing to verify” and proposes a methodology label on every client-facing number. That is the right direction. I want the label attached before the score becomes a target.
Search settings alone can alter the result. In Renown’s 55-prompt test for one B2B brand, ChatGPT named the business in 14.5% of API answers, 23.6% of consumer-app answers, and 38.2% of app answers with search forced on. Same prompts, same week, different conditions. It is a test of one brand, not a forecast for yours. It still makes “we measured ChatGPT” an inadequate description.
Prompt selection matters just as much. Grow and Convert’s argument for topic-based visibility is straightforward: a brand’s visibility depends on the prompts and topics being tested. A headline percentage can rise because a team improves coverage on prompts with little buyer intent. The score moves; the commercial question remains unanswered.
Repetition matters, too. A vendor study of 2,961 runs found ChatGPT and Google AI returned the same brand list less than once in 100 repeats. That finding concerns the exact brand list, not proof that every individual brand’s presence changed each time. Even so, a single-run prompt check is a weak foundation for a performance claim.
My requirement is simple: show what moved, on which prompts, under which conditions. A score without those details is presentation, not diagnosis.
Visibility observations stop before the revenue claim
The next separation is the one that matters most when money changes hands. A citation count is not a revenue measurement. Neither is a mention rate or a recommendation rate.
The available traffic studies show why I would keep these reports separate:
| Dataset | Reported finding | What it does not establish |
|---|---|---|
| Usermaven: 15.3 million sessions across 150 B2B SaaS sites | AI referrals reached 96% of sites but represented 0.88% of total traffic. | That an additional citation produces an additional sale. |
| Marie Haynes’ study of 100 websites over 12 months, covered by Search Engine Land | ChatGPT sessions equaled 0.62% of Google organic sessions; Google organic produced 27 times more conversions than all LLMs combined. | That a higher conversion rate makes AI traffic the larger commercial channel. |
The Haynes study also reports a 13.8% conversion rate for ChatGPT visitors against 3.02% for Google organic. That is worth attention. It does not erase the difference in traffic volume or total conversions. Rate and scale answer different questions, just as they do in the citation study.
Attribution adds another complication. Usermaven reports that more than a third of AI-associated conversions occurred in later sessions, where the returning channel was mostly Direct or Organic search. A referral report therefore does not capture the whole journey either.
None of this makes AI visibility worthless. It means the measurement stops short of the claim “these citations drove this revenue.” A provider making that promise needs evidence beyond a prompt-tracking dashboard.
I want visibility reported as an observation and business performance reported as an outcome. I will not pay for the former as though it proves the latter.
Hire for execution you can audit, not a score you cannot
The procurement questions now follow directly from the data: who was tested, what was asked, how often it ran, which surface produced the answer, and what counted as a hit. Just as important: what will the provider refuse to claim because its measurement cannot see it?
A review of 60 GEO agencies sorts the guarantee language into categories worth keeping apart:
- “Our proprietary visibility score will rise.” I would walk away unless the definition, prompt set, and raw runs are written into the contract.
- “Cited on an agreed share of prompts across named platforms.” Negotiable, provided cited means a link to the business’s own page and the prompt list stays under the client’s control.
- “A guaranteed position, named assistant, or first-citation date.” That promises control over an underlying system the provider does not operate.
I would evaluate groas against the same requirements. The distinction is not that groas gets an exemption from measurement. It is that visibility work sits inside a wider operating model, rather than becoming the product’s finish line.
Specialized models execute campaign, content, bidding, targeting, and budget changes continuously. Actions are logged with their reasoning, so the work can be audited rather than reconstructed from a monthly narrative. A named senior strategist sets direction and guardrails and owns accountability.
The commercial terms are a fixed monthly fee, no setup fee, and month-to-month cancellation. The fee does not rise simply because media spend rises. The performance targets are qualified pipeline, closed-won revenue, ROAS, and cost per acquisition: outcomes that citation studies alone cannot establish.
The published case record reports those downstream measures:
- Search conversions up 89% at 27% lower CPA in month one.
- Hotel bookings up 34% in 14 days at a monthly spend of $40,000–$50,000.
- CPA down 55% in month one.
Those are case outcomes, not evidence that AI citations caused the results or a forecast for another account. That distinction belongs here for exactly the same reason it belongs in a visibility vendor’s report.
Where a pure visibility vendor stops at observations, groas has to keep going: the promise is attributable revenue, not a score. I want the action log, the strategist’s name, and the business outcome—not a broader label pasted over a narrow measurement.
The uncited-business count supports a question about website citations in an uneven panel. It does not support a market-wide invisibility claim, and it cannot establish lost revenue. The buying decision is simpler than the sales language around it: buy execution you can audit, and treat visibility as one diagnostic line inside it.

