“So do we build one plan for ChatGPT and a second one for Claude?”
Ask a room to lay out AI-visibility work and this question arrives within ten minutes, every time.
I stalled on it for the better part of six months. Open any measurement dashboard and the brands ChatGPT names are simply not the brands Claude names. When the outputs disagree that badly, splitting the plan looks like the honest answer.
It also doubles the work. Which is why it’s worth checking whether the disagreement runs all the way down.
Somebody measured that — not the recommendations, but the reasons brands failed to get one.
Should You Build Separate AI Visibility Plans For ChatGPT And Claude?
Not before you diagnose. The brands the two providers recommend diverge about two-thirds of the time, but when both leave a brand out, the diagnosis of why matches 95.1% of the time (Will Jack and colleagues at Unusual.ai, posted May 2026, 215 commercial prompts, 7,763 joint failures).
If neither model is finding you, one diagnosis, run once, carries across both providers. Almost all of that 95.1% is agreement at exactly that stage.
The diagnosis here means sorting your problem into a stage: the models can’t find you, or they find you but leave your name out of the answer, or they name you but stop short of recommending you. The research team sorted every joint failure into those three modes.
215 Commercial Prompts, 7,763 Cases Where Both Providers Left You Out
The team at Unusual.ai went after a practical fork: if your customers use both ChatGPT and Claude, do you optimise per provider (arXiv:2606.26116)?
Fire 215 commercial-context prompts at both providers, line up the recommendations, then pull the 7,763 “joint failures” — the runs where neither provider recommended the brand — and check whether the failure diagnosis agrees.
The divergence gets pinned down first. Cross-provider Jaccard similarity came in at 0.35, while re-running the same prompt against the same provider scores 0.50 to 0.61. The gap between providers is wider than the gap between two runs of the same provider.
The Less Famous You Are, The More The Diagnoses Agree
Agreement overall: 95.1%, confidence interval 94.3% to 95.7%. Split by how well known the brand is, it moves like this.
- Category-leader brands: 81%
- Long-tail and regional brands: 99.6%
The more obscure the brand, the more both providers drop it for the same reason. Turn that around: once you’re well known, your failure modes start to diverge by provider.
Which tells you how long the shortcut lasts. While you’re still building recognition, one diagnosis covers both.
Count only the cases where at least one model found the brand and still dropped it, though, and the diagnoses matched in 67 of 450 (14.9%). If you’re well known and losing at that stage, the paper itself says to diagnose each provider separately.
Different Routes To A Recommendation, The Same Reason For Not Finding You
The two providers reach a recommendation by visibly different roads. Anthropic recommended a brand with no trace of it anywhere in its searches 43-52% of the time; OpenAI did so 8-29% of the time. The authors read that as answering from training-time memory.
One leans more on memory, the other goes looking more often. The authors set the mismatched names and the different roads side by side as two faces of the same picture.
Yet the reason for the stumble is shared. Two people with completely different commutes can still both be late for the same reason: neither one left the house.
If you haven’t left the house, optimising two routes doesn’t help. The paper’s reading is that the more obscure the brand, the less it shows up either on the searchable web or in training data, so both models drop it at the same stage.
How Far This Number Travels
Four limits — who was measured, which markets, what the 95.1% is made of, and who wrote it — and I’d say all four before this goes into a deck.
- Two providers only, OpenAI and Anthropic. Gemini and Perplexity aren’t in it
- The same 215 prompts as the companion study: US, UK and EU markets, 19 industries. No Japanese-language measurement
- Both models file nearly every case as “never found”, so chance alone would produce 93.0% agreement. Corrected for chance (Cohen’s kappa), agreement is 0.29
- All four authors work at Unusual.ai, a vendor in AI brand visibility
That last one bears on the direction of the conclusion. “Your diagnosis is portable” is a comfortable finding for a company that sells diagnosis.
For what it’s worth, it doesn’t contradict the divergence we’ve written about before. What diverges is the output; the classification of causes is the shared part.
Diagnose The Stage Before You Split The Plan
Before you build a provider-by-provider matrix, insert one step. Work out whether you’re losing at discovery — the model never surfaces you — or at selection, where it surfaces you and picks a competitor.
If you’re losing at discovery, that diagnosis carries across OpenAI and Anthropic as it is; for regional brands, 99.6% of diagnoses matched. If you’re losing after being found, the paper’s view is to diagnose each provider separately.
Whether to split the plan is a question worth asking. Just ask it after the diagnosis, not before.
Sources
- Will Jack, Noah Lehman, Keller Maloney, Sarah Xu (Unusual.ai), “Divergent Recommendations, Convergent Diagnoses: Cross-Provider Failure-Mode Convergence in AI Commercial Recommendation”, arXiv:2606.26116v1, posted 22 May 2026, arxiv.org (215 commercial prompts across OpenAI and Anthropic; 7,763 joint failures analysed. Cross-provider Jaccard similarity 0.35 against a same-provider re-run baseline of 0.50-0.61. Failure-mode diagnosis agreed 95.1% [94.3%, 95.7%], rising from 81% for category leaders to 99.6% for long-tail and regional brands. Share of recommendations with no trace of the brand in any search: Anthropic 43-52%, OpenAI 8-29%. Two providers only; prompts shared with the companion study (US/UK/EU markets, 19 industries). The agreement is dominated by cases neither provider found: Cohen’s kappa 0.29, and 14.9% agreement on the 450 cases where at least one provider found the brand; all four authors are employed by Unusual.ai, a vendor of AI brand visibility measurement.)