62% Of AI Brand Recommendations Vanish After One Buyer Question – New Clovion Data

Jul 07, 2026 06:00 PM - 1 week ago 9792

Zahir Hasan didn’t person to show maine his company’s numbers were wrong.

I’d sent Hasan, COO of the Oslo-based investigation patient Clovion AI, a database of methodology questions astir “Surviving the AI Funnel,” Clovion’s caller study of really Claude, ChatGPT, and Gemini urge brands crossed a conversation. Question 10 was routine, the benignant of point you inquire each investigation team. The study says the 3 AI assistants flatly contradict each different connected marque facts 15% of the time, based connected 33 verified contradictions. Was 33 really capable to support a declare astir which exemplary tends to undersell a brand’s features and which tends to oversell them?

Hasan’s reply wasn’t a defense of the number. It was a correction. “The existent number is 330,” he wrote back. “A designer dropped a zero successful layout.” The aforesaid slipped decimal, he said, had besides turned 2,040 brands into “204” connected page 7 of the PDF that I’d been sent successful beforehand of its publication. A revised type is coming retired this week. So, I sewage the corrected figures first.

That’s a unusual measurement to commencement a file astir an AI investigation report, admitting earlier thing other that the draught study had an correction successful it. But it’s the astir honorable measurement in, because the correction says thing the study’s header stats ne'er could. Reading AI answers correctly, whether you’re a marketer trying to fig retired if ChatGPT is recommending your merchandise aliases a interrogator building a study astir it, comes down to catching the decimal constituent earlier you build a strategy connected it.

The Funnel, Recapped

Set the typo speech for a infinitesimal and the underlying investigation holds up. Clovion ran 69,120 multi-turn conversations crossed the 3 assistants successful 36 B2B package and fintech categories, asking an opening mobility for illustration “best CRM tools?” and past a azygous realistic follow-up. Re-asking the aforesaid mobility kept 90% of the recommended database intact. Adding 1 mean purchaser detail, thing arsenic plain arsenic “for a mini team,” kept only 28%. Sixty-two percent of the brands that made the first reply were gone by the 2nd one.

I asked Hasan whether “small team” was cherry-picked to nutrient that drop. It wasn’t. His squad besides tested “for a ample enterprise” and sewage almost identical churn, astir 72% either way, against astir 10% erstwhile the mobility was simply repeated. The database isn’t unstable. It’s responsive, and mostly to whether the exemplary has decided who a marque is really for.

That’s the portion worthy sitting pinch if you do SEO aliases marque strategy for a living. Being named successful an AI reply is not the aforesaid point arsenic being trusted by it. A exemplary that puts you successful its first CRM database tin still trim you the infinitesimal a purchaser gets specific, and Clovion’s information says that happens astir of the time, not immoderate of the time.

The Correction Changes the Shape of the Smallest, Most-Cited Number

Here’s wherever the fixed decimal really matters for really you should publication this study. The aged figure, 33 verified contradictions, was mini capable that immoderate per-model declare built connected it was opinionated connected bladed ice. Corrected, it’s 330, and the per-model breakdown Hasan shared is acold much telling than the aggregate 15% fig the draught study leads with: Claude underclaims a brand’s ain features 160 times against 10 overclaims. ChatGPT underclaims 70 times and ne'er overclaims. Gemini runs the different way, overclaiming 80 times against 30 underclaims.

Hasan’s moving theory, drawn from a separate, not-yet-published Clovion study connected wherever each exemplary sources its answers, is that Gemini leans much heavy connected trading worldly and video, truthful it tends to in installments a marque pinch immoderate it’s hyping. Claude and ChatGPT thin much connected archiving and merchandise pages, picture the halfway merchandise accurately, and hedge toward “doesn’t person it” erstwhile a newer characteristic isn’t good documented. If that holds up nether the study Clovion hasn’t released yet, it intends the guidance of an AI assistant’s correction astir your merchandise is simply a usability of what benignant of contented you’ve put successful beforehand of it, and wherever that contented lives.

I’ve spent much than 20 years telling clients that ranking good and being described accurately are 2 different problems. This is the clearest grounds I’ve seen that they’re now the aforesaid problem, playing retired wrong a azygous conversation, and that the hole depends connected which adjunct is doing the misdescribing.

Why Nobody Catches the Missing Zero

Frederick Vallaeys has a communicative successful his book “The AI-Amplified Marketer” that explains precisely why a dropped decimal survives each the measurement to publication. An automated study erstwhile flagged “great performance” connected a keyword because its costs per acquisition was moving overmuch higher than the target. Somewhere successful the system, precocious had gotten swapped for good, erstwhile a precocious CPA is bad news, not bully news. Anyone skimming the summary would person nodded along, because the condemnation publication smoothly moreover though its meaning had flipped.

Vallaeys ties this to investigation connected predictive processing, the thought that fluent readers aren’t decoding each word, they’re predicting what comes adjacent based connected discourse and moving on. That’s really “teh” sounds arsenic “the” and a missing “not” slides correct past you. As Vallaeys puts it, our intelligence exemplary of the condemnation overrules the matter successful beforehand of us. A confident, well-formatted PDF is the easiest spot successful the world for that to happen, and a dropped zero successful a layout record is simply a overmuch smaller, overmuch much forgivable type of the aforesaid failure.

It’s besides why the hole isn’t “trust the study less.” It’s “keep a quality aviator successful the loop who checks the number alternatively of the vibe of the paragraph astir it.” Thirty-three contradictions and 330 contradictions don’t conscionable disagree by a facet of ten. They support wholly different assurance levels astir whether a per-model shape is real. Two 100 4 brands and 2,040 brands aren’t the aforesaid study. If Clovion hadn’t caught it, and if I hadn’t asked, the smaller, shakier numbers would person kept circulating arsenic fact, cited by precisely the benignant of waste and acquisition property that’s expected to drawback this.

What Clovion Isn’t Claiming, and Why That’s the Honest Part

The study is observant to opportunity the nexus betwixt really a exemplary perceives your fresh and whether it recommends you is “a strong, accordant coupling, not a proven causal law.” I pushed Hasan connected what a existent causal trial would look like. His answer: alteration 1 thing, a brand’s nationalist positioning content, time off everything other alone, and spot whether the models’ behaviour moves comparative to brands cipher touched. Clovion hasn’t tally that trial yet. He besides conceded the much uncomfortable anticipation directly, that a brand’s existent real-world positioning is astir apt driving some really the exemplary describes it and whether it gets recommended, which would make positioning the existent lever and the model’s “perception” conscionable a symptom, not a cause.

That’s an unusually candid reply from a institution trading AI visibility monitoring, and it’s precisely why I spot the remainder of what Hasan told me. He besides had nary information connected really accelerated an AI’s cognition of a marque shifts aft that marque changes its ain content. “We didn’t do a before-and-after test,” he said. “Treat it arsenic worthy testing, not guaranteed successful X weeks.” Anyone telling you they tin committedness a circumstantial timeline for moving Claude’s aliases Gemini’s sentiment of your marque is guessing, by Clovion’s ain admission.

What To Actually Do About It

There are 3 things that you should do, based connected what Hasan told maine and what the corrected information supports.

First, way the full conversation, not the first answer. If you’re monitoring AI visibility pinch a single-prompt check, you’re measuring the apical of a chimney that loses 62% of its contents 1 condemnation later. Build your monitoring astir the follow-up questions your existent buyers really ask.

Second, hole the assistants 1 astatine a time, successful order. Hasan was nonstop that a azygous contented alteration won’t move each 3 models astatine once, because they propulsion from different sources. His suggested order: correct level actual errors first, since those are inexpensive wins, past spell aft the segment-fit combinations that matter astir to your pipeline, checking each adjunct crossed respective runs alternatively than trusting immoderate azygous answer.

Third, don’t mention a stat you haven’t traced to its source, including this one. Clovion’s ain study needed a correction connected its astir technical, astir citable number. Before you build a column, a customer deck, aliases a contented little astir immoderate AI investigation percentage, inquire wherever the underlying count came from and whether anyone’s checked the mathematics since it near the creation software.

I’ve watched SEO spell done a fewer of these moments, from Panda to mobile-first indexing to the slow bleed of zero-click search. Each 1 rewarded the practitioners who checked the superior root alternatively of repeating the header number. AI visibility is shaping up the aforesaid way. The brands that triumph the disappearing enactment Clovion documented won’t beryllium the ones pinch the champion property merchandise astir their AI Overviews strategy. They’ll beryllium the ones who publication the study intimately capable to inquire what a “33” really meant, and who support asking that mobility aft this one.

Zahir Hasan is COO of Clovion AI, based successful Oslo, Norway. Clovion’s corrected type of “Surviving the AI Funnel,” reflecting the figures successful this column, is expected this week.