Why Does My Brand Disappear From AI Answers?

Run a query. Get five brands. Rephrase it — “top” instead of “best,” a location added, a price constraint — and two of them disappear. The AI visibility industry has spent two years selling this as an optimization problem: appear more consistently, build more entity signals, publish more content. The research says they’ve aimed at the wrong target. The instability you’re measuring is real, but it’s a symptom. The actual problem sits one level down — and it’s the thing that determines whether any of the standard moves actually work.

Your score isn’t a ranking

A study published in May 2026 ran a controlled experiment on AI product recommendations.^[1] The setup was simple: take a query, paraphrase it, compare the brand sets returned. A cosmetic rewording — changing “best running shoes for long distance” to “top long-distance running shoes” — produced recommendation sets with a Jaccard similarity of 0.288.

Jaccard is just overlap — 1.0 means identical lists, 0 means nothing shared. So 0.288 means roughly a quarter of the brands survived the rewording, and three-quarters didn’t. Against a rerun baseline of 0.50–0.61 (same prompt submitted twice, same model), the drop from a cosmetic rewording is substantial. Add a constraint — a location, a price ceiling, a specific use case — and overlap falls to 0.135. That means 86.5% of the brand set is non-overlapping across variants of what is, in intent, the same question.

If you run any kind of AI visibility measurement program, these numbers carry a direct implication: a single query submitted once is not a measurement. It’s one sample from a distribution whose shape you don’t yet know. The score on your dashboard may look like a ranking. It isn’t.

The instinct is to run more queries and average the results — reduce the variance, stabilize the signal. That instinct is wrong. Not because the instability isn’t real, but because it mistakes the symptom for the disease.

The noise isn’t the story

Before getting to what the instability actually means, it’s worth establishing what it doesn’t — because the GEO and AI visibility industry is currently building a large commercial apparatus on the wrong premise.

Independent statistical analysis of AI citation distributions shows that AI recommendation outputs follow a power-law distribution: a small number of brands capture a disproportionate share of recommendations, most others tail off rapidly, and run-to-run variability is high.^[7] The practical consequence: most of what looks like a meaningful difference in brand visibility between query runs falls inside the noise floor of a single measurement.

Think of it this way. If you run 50 queries and your brand appears in 30, you don’t know your AI visibility. You know one sample. The underlying distribution could be consistent at 60%, or it could be bimodal — appearing in almost every query under some phrasings and almost none under others. Those are completely different strategic situations. A point estimate cannot tell you which one you’re in.

The AI visibility score your vendor is selling you is one reading from a system you haven’t mapped — not a fixed value, not a ranking. The industry is selling certainty about an inherently uncertain phenomenon. That’s the false lead. And it matters because the right response to this kind of uncertainty isn’t more queries — it’s understanding the structure underneath the noise.

Models that disagree still fail the same way

Here is what’s actually interesting.

A cross-provider study published in June 2026 ran the same recommendation queries across paid API configurations of OpenAI and Anthropic models — GPT-5.4-mini, GPT-5.4 flagship, Claude Sonnet 4.6, and Claude Opus 4.6, all with native web search enabled.^[3] The models disagreed on brand recommendations roughly two-thirds of the time — a cross-provider Jaccard of 0.35. On its own that number is unremarkable: two different providers should disagree more than one model disagrees with its own rerun. It matters here as backdrop — this much surface disagreement is what makes the next finding strange.

But then the researchers looked at what happened when both providers’ models failed to recommend a brand. To understand this finding, you need to know how AI recommendation systems work in two stages. First, retrieval: the model assembles a candidate pool — a shortlist of brands it considers relevant to the query. Second, selection: from that pool, it picks which brands to actually recommend in the final answer. Most AI visibility optimization work targets retrieval — getting your brand into the candidate pool. The research says many of the interesting failures happen at selection — being in the pool but not getting picked.

When the OpenAI and Anthropic models both omitted the same brand, the same underlying failure mode was identified 95.1% of the time, across three categories: discoverability (the model couldn’t reliably retrieve or identify the brand — it failed before the question was even answered), compellingness (the brand appeared in retrieval but wasn’t selected — it made the pool and then lost the argument), and positioning (the brand appeared and was sometimes selected, but shifted dramatically based on how the AI framed the user asking the question).

Read that figure carefully — including what it is not. The failure modes describe what the model actually did — did the brand enter the candidate pool, did it survive selection, did it swing with the imagined user — as sorted by the researchers, under a single rubric. The models are not explaining themselves. With that said: two providers that disagree two-thirds of the time on which brands to recommend show the same failure pattern 95% of the time when a brand goes missing.

This is not noise. This is structure. The instability in what the models recommend is real, but the failure pattern underneath it is shared and stable. The same brand gets excluded for the same reason across systems that otherwise behave very differently.

A companion audit across 37,000 runs adds texture to this.^[2] The instability isn’t uniformly distributed. L1 category leaders — the dominant brands in any category — appear in nearly every retrieval but capture only 25–41% of final recommendation slots. Being retrieved constantly is not the same as being recommended. L4 and L5 specialists and regional players are entirely absent from 48–52% of runs. And mid-market brands sit in the most volatile position: persona conditioning caused them to swap up to 75% of their recommendation set, while category leaders remained roughly 80% consistent across personas.^[4]

Persona conditioning means prefixing the same query with a user context — “as a budget-conscious buyer, recommend running shoes” versus “as a professional athlete, recommend running shoes.” Same product category, different imagined user. For category leaders, it barely moves the needle. For mid-market brands, it reshuffles almost the entire recommendation set.

This stratification is counterintuitive. Conventional wisdom says larger brands benefit from AI visibility because they appear more consistently. The finding is that larger brands have a different problem: they appear in the retrieval pool almost always, but convert that retrieval into a final recommendation at a lower rate than expected. Specialists who are absent entirely have a different problem still. Mid-market brands on the persona fault line have a third problem that’s unrelated to either. Standard AI visibility measurement — a single query, a point estimate — collapses all three into one undifferentiated score, and tells you nothing useful about which problem you actually have.

Why you can’t average your way out

The standard response to LLM unreliability is to reduce variance through engineering: run the model more times and average the results, lower the temperature to make outputs more deterministic, add retrieval-augmented generation to ground the model in facts.

None of it works — and recent work explains precisely why.

Retrieval-augmented generation (RAG) is the technique where a model, before answering, looks up relevant information from a knowledge base or database. It’s the architecture behind most enterprise AI search, and the standard recommendation in the AI visibility playbook: ensure your brand is in the retrieval corpus, make sure your entity signals are clear, make sure the model can find you. The assumption is that grounding the model in retrieved facts makes it more consistent and accurate.

Research published in June 2026 found that when RAG systems got an answer wrong, the error often would not budge: in 42% of cases for one common architecture and 59% for another, the system gave the same answer on all five samples.^[5] Not scattered wrong answers — the identical wrong answer, five times out of five. Standard uncertainty-detection methods, which work by looking for inconsistency across multiple runs, are completely blind to this failure mode. The model looks stable. It is stably wrong.

Earlier work reached the same conclusion from the training side: neither retrieval nor temperature tuning can guarantee stable answers to equivalent prompts, and the only proposed cure — retraining the model itself — is an early-stage research idea, not something you can buy.^[8] One honest boundary on this evidence: it comes from studies of factual question-answering, not brand recommendation, so it does not prove the exclusions we saw earlier are locked in the same way. What it proves is narrower but enough: a stable answer is not a correct answer, and rerunning the same prompt cannot expose a failure the model commits identically every time.

What this means in practice: running more queries and averaging the results doesn’t change what the model does. In the cases that matter most — the stably wrong ones — it hides a fixed answer behind a false sense of measurement. Picture the monthly report: your visibility moved from 31% to 34%, and someone is about to get budget to keep the momentum going. If the number underneath is a fixed exclusion, the chart is congratulating variance. You are averaging signal you can’t see into noise that looks like data. The way out is not more reruns of the same prompt — it is variation across the dimensions the failure actually lives in: phrasings, personas, models. Under that kind of sampling even a stable exclusion shows itself, because consistent absence across every condition is itself a diagnostic. That is the sampling the final section prescribes.

What this evidence can’t prove yet

This is a good place to slow down. The case above rests on an evidence base with real limitations.

The core research here is almost entirely the work of a single team at Unusual.ai, published across several papers in May and June 2026, all unreviewed arXiv preprints at time of writing.^[1][2][3][4] The 95.1% failure-mode convergence figure — perhaps the most striking finding — has not been independently replicated. The models tested were GPT-5.4-mini and Claude Sonnet 4.6 in the primary paper, expanding to GPT-5.4 flagship and Claude Opus 4.6 in the cross-provider work — all via paid API tiers with native web search enabled, not the consumer chat interfaces. No open-source models, no Gemini. The domain was commercial product recommendation; whether the findings generalize to services, B2B, or other categories is untested.

The measuring stick itself is also under challenge. The same pair of outputs has scored 0.75 on one similarity metric and 0.10 on another — a 65-point swing from metric choice alone.^[9] Jaccard specifically can hand identical scores to lists that are nearly identical and to lists that share almost nothing.^[10] And simply changing how outputs were judged has been shown to shrink apparent instability by up to 55-fold on the same models — some of the measured chaos may live in the ruler, not the system.^[11] That reaches every Jaccard figure in this essay: the 0.288 and 0.135 that opened it, the 0.35 between providers. Hold their magnitudes loosely; nothing that follows rests on them.

These are real critiques. Hold them.

But notice what they do and don’t establish. The metric critiques challenge the size of the instability numbers — if the true Jaccard were 0.40 rather than 0.288, the finding is less dramatic. They don’t touch the shared-failure finding, because that one rests on a different observation: not how much the models disagree, but that when they fail, they fail the same way. The limits of a similarity score have no bearing on that.

The one-team problem is more troubling, and the sharpest version of the worry is not fabrication. It is that the 95.1% could be partly an artifact of who did the sorting: one team, one rubric, three categories, applied to both providers’ outputs — and a blind spot in the sample, since brands both providers omit will skew obscure, and for obscure brands discoverability is the easy call. What limits the worry without dissolving it: the categories describe things you can observe — in the pool or not, picked or not, persona-sensitive or not — which leaves less room to manufacture agreement than a set of judgment calls would. An independent recount would settle it. It hasn’t happened. The independent corroborations that do exist — the same wording-brittleness showing up in plain factual question-answering on standard benchmarks^[6], and the statistical noise-floor analysis reaching the same skewed-distribution conclusions from an entirely separate starting point^[7] — cover the most important part of the gap: whether any of this holds beyond one team’s experiments.

The skeptic is right that the magnitude is uncertain and that independent replication hasn’t happened. The skeptic is not right that this voids the strategic implication. The strategic implication doesn’t depend on the precise Jaccard number — it depends on the three failure modes. And be precise about what those are: a sorting of symptoms you can observe — where in the pipeline a brand falls out. That each symptom responds to a different fix is the framework’s working bet, not its proven result. No direct rebuttal of it exists anywhere in the research as of this writing; no independent confirmation does either.

The industry has also produced its own counter-study, and it deserves a different kind of scrutiny. In June 2026, the chief product and marketing officer of Peec AI — a platform that sells AI visibility tracking — published a sponsored analysis of 37,804 AI responses across five engines, concluding that prompt wording matters less than intent: brand mentions hold steady as long as the core intention stays the same, with visibility dropping by half only once prompt similarity collapses toward the bottom of the scale.^[12] On its face, a direct rebuttal of the paraphrase-brittleness finding.

It is not a rebuttal. It is the same phenomenon measured with the instrument this essay has spent four sections arguing against. The study aggregates brand mention rates across tens of thousands of responses — a mean, exactly the point estimate that smooths away the per-run churn the Jaccard analysis exposes. A stable average is not evidence against an unstable distribution — averaging makes any distribution look calm, which is why the headline number cannot settle the question in either direction.

The intent measure inherits the same blindness: short-prompt embeddings score “best running shoes” and “best running shoes under $200” as near-identical in meaning, which quietly files the constraint paraphrase — the exact case that drove overlap to 0.135 — under “same intent, stable.” And the study’s own secondary findings reproduce the structure it claims to rebut: keyword-style prompts lifted brand visibility by up to 25% over conversational phrasing, ranking-format prompts surfaced roughly 20% more brands than open-ended questions, middle-of-funnel queries were the most wording-sensitive, and branded bottom-of-funnel queries showed what the author himself labels false stability. Those differences survive the critique that fells the headline, because the objection is to a mean posing as a measurement, not to comparisons between conditions: averaging blurs differences, it does not invent them. Framing effects of that size, with intent held constant, are this essay’s opening finding wearing the rebuttal’s byline.

That the reassuring headline sits on top of that data — sponsored, published by a vendor whose product is the point estimate — is not a coincidence. It is the point-estimate industry writing its own defense. Read the sub-findings, not the headline. And note what the study never measures: why a brand is absent. The failure modes go untouched.

Which problem do you actually have?

The thesis, restated: AI recommendations don’t have noise. They have structure. The instability practitioners are racing to smooth over is, at its core, a set of shared, stable, structural exclusion decisions — and the standard optimization toolkit cannot reach them.

That reframe has a direct implication for what GEO practitioners and their clients should actually be doing.

First: stop measuring with single-run point estimates. A single query submitted once to a single model tells you almost nothing diagnostic. What you need is the distribution — the same intent run across enough prompt variants, personas, and models to reveal where you usually land and identify the conditions under which you are and are not included. Without the distribution, you don’t know which failure mode you’re in. Without that, any fix is a guess.

Second: once you have a distribution, the question shifts from “how do we appear more often?” to “which failure mode are we in?” The answer determines everything downstream.

If your brand is consistently retrieved but converts to only a fraction of final recommendation slots, you have a compellingness problem. More content, stronger entity signals, better schema markup — none of these fix compellingness. You’re already being found. Something in how you’re described, differentiated, or compared at the moment of recommendation is losing the argument. The optimization target is the selection step itself: how the AI describes you against the alternatives at the moment it makes the recommendation.

If your recommendations shift dramatically based on how the question is framed or who the AI imagines is asking, you have a positioning problem. The challenge isn’t what you say — it’s who the AI thinks you’re for. Persona-specific content signals and clearer audience anchoring may matter more than generic authority-building. Thought leadership aimed at the wrong imagined user compounds the misalignment. (A note on definitions: the research measures positioning by swapping personas; the wording-driven churn that opened this essay is a close cousin, measured separately — treating them as one family is this essay’s call, not the papers’.)

If you’re absent from half or more of all runs — even for queries where you should clearly appear — you have a discoverability problem. The retrieval corpus doesn’t confidently know who you are, what category you belong to, or why you’re relevant. Entity definition comes first: consistent naming, structured schema, clear category signals. Content investment before entity clarity is content the model still can’t retrieve or classify. One thing the three modes don’t untangle: absence can be rooted in the live retrieval layer or in the training corpus itself — a slower channel, built through entity co-occurrence, that schema work barely reaches.^[13] The research this framework rests on does not separate the two.

These are not the same problem. And the fixes for one are counterproductive when applied to another.

The insight this body of research puts on the table is that AI visibility is not a ranking you optimize by climbing. It is a diagnosis you survive by naming correctly. The wrong diagnosis wastes time on the wrong lever while the real exclusion goes unaddressed.

That is what stability scores cannot tell you. And it is what the distribution, and the failure modes behind it, can.

What changes Monday morning

If your AI visibility product delivers a point estimate, you are selling a sample from a distribution your client cannot interpret. That is not a technical nicety — it is a structural problem with how AI visibility is currently measured, reported, and sold. A brand manager who receives a monthly “AI share of voice” score cannot tell whether their brand has a discoverability problem, a compellingness problem, or a positioning problem. Those three problems have different causes, different costs, and different fixes. A single number collapses them into noise.

The measurement shift. Distribution-based measurement costs more than a point estimate — a few dozen runs where there was one — but the spend is API calls, not headcount, and it is trivial next to the cost of shipping the wrong fix. The minimum viable version: take one core intent, run it across five to ten phrasings, two to three personas, and two search-enabled API-tier models. That produces enough spread to see whether your brand appears consistently or sporadically, whether persona conditioning reshuffles your set, and whether cross-model behavior is coherent or divergent. You now have a distribution. The shape of that distribution is your diagnostic.

This is different from running more queries to average a score. The goal is not a more stable number — it is a distribution with enough variance to reveal which failure mode you are in. If your results are tightly clustered across all phrasings and personas, look at compellingness next. If your results shift dramatically by persona, you have a positioning problem. If you are absent across most conditions, you have a discoverability problem. These diagnoses are available from the distribution. They are invisible from the point estimate.

One practical note on seeing the pool. The two API tiers this measurement runs on can expose their sources, but unevenly: Anthropic’s responses list every result the search retrieved, with the cited ones marked separately, so retrieved-but-never-cited is readable out of the box; OpenAI’s show only the cited URLs unless you explicitly ask the API for the full list of sources it consulted — skip that setting and the compellingness signal is invisible on that leg. And what you see is sources, not brands: a page that made the grounding but never got cited is a signal pointing at compellingness, a proxy for the brand-level failure rather than the thing itself. If all you have is chat interfaces, the fallback is rank position — Petrovic reports that within Google’s grounding, a result near the top of traditional search tends to sit near the top of the cited set.^[16] That is one observation about one retrieval stack; treat it as a lead worth testing on the Bing- and Brave-backed retrieval the other models use, not a law.

The measurement half of this shift is already standard practice at the sharp end of the industry. Charles Floate maps AI Overviews with thirty-plus captures per query, sorted into a stable core that appears every time and a volatile layer that comes and goes — and reports that the pool of citable sources holds remarkably steady even as the generated answers churn, which is this essay’s structure-not-noise claim observed from the field. The scopes differ, and both hold: his stability is per query, across repeated captures of one wording; the audit’s 48–52% absence rates are across phrasings and personas.^[14] AI Overviews monitoring more broadly has standardised on repeated sampling — multiple runs per query per engine, tracked over weeks — because a single run is treated as noise, not signal.^[15] And Dan Petrovic probes models the way a statistician would: mention frequency per hundred probes, average ordinal position in the answer text, and, the sharpest cut, what his dashboard calls citation share tracked separately from mention share — presence in the model’s grounding pool versus being named in the generated answer, because a brand can be either without the other.^[16]

Be precise about what this does and does not establish. Repeated sampling is measurement hygiene, not proof of the three failure modes — the industry was sampling distributions before these papers existed. But the pieces of the diagnosis are appearing in the wild: Floate’s stable pool localizes the churn downstream of retrieval, and Petrovic’s citation-strong, mention-weak brands — in the pool, never named — are the compellingness symptom sitting on a live dashboard. What no tooling yet ships is the classification itself: the sorting of absence into discoverability, compellingness, or positioning that turns a well-sampled distribution into something more than a better score.

The deliverable changes. The output of a proper AI visibility audit is not a rank. It is a failure mode diagnosis. A score tells a brand manager they appear in 34% of AI responses. A diagnosis tells them they appear in 62% of category-leader queries, 12% of budget-buyer queries, and 8% of regional queries — and that the cross-model convergence on omissions is consistent, pointing to a positioning failure rather than a discoverability gap. Those are different documents. The second is actionable. The first is not.

Agencies that can produce the second document are selling something structurally different from agencies that cannot. That gap is currently invisible in the market because most AI visibility products look similar at the pitch level. It will not be invisible when clients start asking why the fixes they bought did not move the distribution.

Three fixes to stop selling. The research implies a meaningful share of current AI visibility spend is allocated to the wrong intervention. If the three-mode sorting is wrong, the diagnosis sends the wrong fix — but the method carries its own correction: you measured the distribution before the fix, so you will see whether the fix moves it. A point estimate never offered that feedback loop.

Content volume does not fix a compellingness problem. If your brand is already being retrieved and failing at the selection step, more content is more of the thing the model already found insufficient. The model knows you exist. It is choosing not to recommend you anyway. That is a framing and differentiation problem at the recommendation moment — not a content-volume problem.

Schema and entity cleanup do not fix a positioning problem. If your brand shifts dramatically by persona, your entity structure is probably not the issue. The issue is that the model’s representation of who your brand is for is inconsistent or wrong. Structured data cannot fix that directly. Persona-anchored content can.

Thought leadership does not fix a discoverability problem. If the model cannot reliably retrieve or identify your brand, publishing authority content compounds an identity problem — the model adds content it cannot attribute clearly to an entity it does not have a confident handle on. Entity definition comes first: consistent naming, category signals, and structured schema before any content program.

The action sequence. Measure the distribution first. Diagnose the failure mode second. Apply the right fix third.

The current market runs this in reverse: apply a fix (content, schema, thought leadership), measure whether the score moved, attribute causation. That approach cannot distinguish between a fix that worked and a fix that coincided with random variance. The distribution makes that distinction visible.

A brand manager who receives this analysis can do something the current generation of AI visibility reports does not enable: forward it to their agency with a specific question. Not “why is our score low?” but “we have a compellingness failure — what are you changing about how we’re framed at the selection step?”

That is a different conversation. It is also a harder one to answer with generic deliverables. That difficulty is the point.

References

1. Jack, W., Lehman, N., Maloney, K., & Xu, S. (2026). Paraphrase brittleness in product recommendation: How query phrasing destabilizes AI-generated brand sets. arXiv. https://arxiv.org/abs/2605.27440
2. Jack, W., Lehman, N., Maloney, K., & Xu, S. (2026). Prominence-stratified failure modes in retrieval-augmented commercial recommendation: A 37,000-run audit. arXiv. https://arxiv.org/abs/2605.27439
3. Jack, W., Lehman, N., Maloney, K., & Xu, S. (2026). Divergent recommendations, convergent diagnoses. arXiv. https://arxiv.org/abs/2606.26116
4. Jack, W., Lehman, N., Maloney, K., & Xu, S. (2026). Persona conditioning of brand recommendations in retrieval-augmented commercial chat. arXiv. https://arxiv.org/abs/2605.30207
5. Julka, S. (2026). When confidence takes the wrong path: Diagnosing retrieval-state lock-in in RAG. arXiv. https://arxiv.org/abs/2606.22728
6. arXiv. (2025). LLMs show surface-form brittleness under paraphrase stress tests. NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle. https://arxiv.org/abs/2510.08616
7. Sielinski, R. (2026). Quantifying uncertainty in AI visibility: A statistical framework for generative search measurement. arXiv. https://arxiv.org/abs/2603.08924
8. Prabhune, A., et al. (2025). Information-consistent language model recommendations through group relative policy optimization. arXiv. https://arxiv.org/abs/2512.12858
9. Burger, B., Walter, M., & Le, H. (2024). Metric sensitivity in XAI explanation stability evaluation. arXiv. https://arxiv.org/abs/2406.15839
10. Petkovic, M., & Skrlj, B. (2020). On the limitations of Jaccard similarity for ranked list comparison. Applied Soft Computing. Elsevier. https://arxiv.org/abs/2008.02216
11. arXiv. (2025). Flaw or artifact? Rethinking prompt sensitivity in evaluating LLMs. arXiv. https://arxiv.org/abs/2509.01790
12. Landwehr, M. (2026). Prompt tracking: Does prompt variance % impact brand mentions? Search Engine Journal (sponsored by Peec AI). https://www.searchenginejournal.com/ai-prompt-intent-keywords-peec-spa/576201/
13. Floate, C. (2026). Every AI answer your customers see is generated through five layers [LinkedIn post]. LinkedIn. https://www.linkedin.com/posts/charlesfloate_every-ai-answer-your-customers-see-is-generated-activity-7480187965776961536-ivsi
14. Floate, C. (2026, July 11). I reverse engineered 1000s of AI Overviews: Here’s exactly how to get Google to cite YOU [Article]. X. https://x.com/Charles_SEO/status/2075894365348860308
15. Geneo. (2026). The ultimate guide to AI search volatility tracking. Geneo. https://geneo.app/blog/ultimate-guide-ai-search-volatility-tracking/
16. Petrovic, D. (2026). What is AI SEO [Webinar]. Edge of Search Future Proof webinar series. YouTube. https://www.youtube.com/watch?v=QlK_fvAw_9E