AI Recommends You Sometimes but Not Others — Here’s What Triggers the Difference

The intermittent-visibility problem

You asked ChatGPT to recommend products in your category. Your brand appeared. You asked again the next day and it was gone — replaced by a competitor you have never heard of. You asked a third time and you were back, but listed last instead of first. Nothing changed on your end. No new press. No website update. No pricing change. Yet AI’s answer about you shifted completely between queries.

This is not a glitch. It is a signal. Your brand sits at the boundary of what AI considers confidently recommendable for this topic. Sometimes the probability tips in your favor. Sometimes it does not. The question worth asking is not “why did it change?” but “what determines the probability that we appear at all?”

Why the pattern is not random

Large language models are probability engines. Every response is generated by sampling from a distribution of likely next tokens. This makes individual outputs nondeterministic — the same prompt produces different answers every time. SparkToro’s research across 2,961 queries confirmed that AI recommendation lists repeat less than 1% of the time.

But nondeterministic does not mean patternless. Underneath the per-query variation, there is a stable probability distribution. Some brands show up in 90% of relevant queries. Others show up in 15%. The per-query randomness makes the pattern invisible when you spot-check, but it becomes clear when you measure at scale.

Think of it this way: a coin flip is random, but a loaded coin still lands heads more often. The “loading” of AI’s coin — the factors that determine how often your brand lands in the answer — follows predictable rules. The SparkToro study found that in constrained categories like healthcare, top brands appeared in 55% to 97% of responses. In broader categories, the field was more scattered. Whether your brand sits at 90% or 15% is not luck. It is a function of how AI evaluates you against specific criteria.

What triggers inclusion

When AI decides whether to include your brand in a recommendation, it is not checking a ranking list. It is evaluating confidence — how sure it is that recommending you is a correct, well-supported answer to the question being asked. Multiple factors compound to build or undermine that confidence.

Corroboration from independent sources

The single strongest predictor of consistent AI inclusion is what researchers call the corroboration threshold: the point at which 2–3 independent, authoritative sources all confirm the same claims about your brand. Below the threshold, AI hedges. It says “claims to be” instead of “is.” It includes you in some outputs but not others. Above the threshold, confidence locks in and your mention rate stabilizes.

This is multiplicative, not additive. Search Engine Land documented a cascading confidence model: if each processing stage passes 90% confidence across 10 stages, total confidence drops to 35%. A single weak link at 50% drops the total to 19%. One gap in corroboration can drag your entire inclusion probability down.

Positioning consistency

AI cross-references your website, marketplace listings, review sites, LinkedIn page, and other public data sources. If those sources describe your brand consistently — same category, same core claims, same positioning — AI’s confidence increases. Brands with consistent editorial mentions achieved AI recommendation rates 89% higher than brands with scattered, contradictory positioning across their web presence.

Category specificity of the query

Narrow queries produce more consistent results. When someone asks about “cloud computing providers,” the field is small and well-defined, so top brands appear in nearly every response. When someone asks about “best project management tools,” the category is broad enough that AI rotates through dozens of options. Your mention rate on broad queries depends on how strongly AI associates you with the specific sub-need the user described.

Content freshness and volume

AI models pull from training data and, increasingly, from live retrieval. Brands that produce consistent, category-relevant content create more touchpoints for AI to find and corroborate. Data from Brandi AI found that brands producing 12+ optimized pieces of content saw up to 200x faster visibility gains compared to those producing four or fewer. This is not about volume for its own sake — it is about giving AI more evidence to build confidence.

What triggers exclusion

Understanding why AI skips you is as important as understanding why it includes you. Several factors actively suppress inclusion, even when other signals are strong.

Positioning contradictions

If your homepage says “enterprise security platform,” your G2 listing says “compliance automation tool,” and a press article calls you a “cybersecurity startup,” AI cannot confidently categorize you. This ambiguity reduces the likelihood of inclusion in any specific answer. The model is not confused — it is rationally hedging because your own sources disagree about what you are.

Vocabulary mismatch

Your brand uses internal terminology. Buyers use different language when asking AI. If your website talks about “unified observability” but buyers ask for “server monitoring tools,” AI may not connect the two. The mismatch does not make you invisible everywhere, but it makes your inclusion dependent on whether the specific phrasing of each query happens to overlap with your content — which is exactly the kind of intermittent visibility you are experiencing.

Negative sentiment signals

AI models weigh sentiment. If your brand has significant negative coverage — poor reviews, unresolved PR issues, or customer complaints dominating community discussions — positive mentions alone will not override those signals. Brands with low review scores or viral controversies can find themselves omitted from AI-generated answers even when they otherwise meet the inclusion criteria.

Blocked AI crawlers

Many companies have robots.txt configurations that inadvertently block AI crawlers like GPTBot, ClaudeBot, or PerplexityBot. If AI cannot retrieve your content during live-search-augmented responses, your visibility in those sessions drops to whatever the model remembers from training data alone — which may be outdated or insufficient. This creates a pattern where you appear in some queries (when the model answers from parametric memory) but not others (when it relies on retrieval).

The three knowledge layers AI draws from

AI does not retrieve answers from a single source. It synthesizes across three distinct knowledge representations, each with its own strengths and failure modes. Understanding these layers explains why your brand might appear in some types of queries but not others.

Entity graph (structured knowledge)

This is AI’s factual database — structured information about companies, products, people, and their relationships. If your brand has a well-defined entity in knowledge graphs (think Wikipedia, Wikidata, Crunchbase, or well-structured schema markup on your own site), AI can confidently identify what you are. Brands with weak entity presence get treated as ambiguous — AI is less willing to recommend something it cannot clearly categorize.

Document graph (indexed web content)

When AI platforms use retrieval-augmented generation — pulling live web results to inform their answers — they draw from an index of web content. Your placement here depends on whether authoritative sites have published content that mentions your brand in relevant contexts. Review sites, industry publications, comparison articles, and community discussions all contribute. If the document graph contains strong, recent, consistent mentions of your brand for a topic, you appear more often in retrieval-augmented responses.

Concept graph (learned associations)

This is the “gut feel” layer — associations the model learned during training from the massive corpus of text it ingested. If your brand was frequently mentioned in training data alongside your category, that association persists even without retrieval. But this layer is also the fuzziest. It reflects whatever the training data contained, including outdated information, and it cannot be directly updated outside of model retraining.

Brands present across all three layers receive disproportionate visibility. The layers reinforce each other: the entity graph confirms what you are, the document graph provides recent evidence, and the concept graph provides the learned association. When one layer is missing, your inclusion becomes inconsistent — you appear when the query happens to lean on a layer where you are strong, and you disappear when it leans on a layer where you are weak.

How to measure your actual inclusion rate

You cannot see the pattern by spot-checking. When every query produces a different answer, a handful of manual tests tells you almost nothing about your real inclusion rate.

Meaningful measurement requires:

• Volume: SparkToro’s methodology used 60–100 queries per prompt per platform. At fewer than 50 queries, variance is too high to distinguish real patterns from noise. A brand that appears in 3 out of 5 queries could have a true inclusion rate anywhere from 20% to 80%.
• Prompt diversity: Buyers phrase the same need differently. “Best CRM for startups” and “what CRM should a small team use” are functionally identical but produce different AI outputs. You need to test across the full range of phrasings your buyers actually use.
• Platform coverage: A brand might score 75% inclusion on ChatGPT and 30% on Perplexity for the same queries. Each platform uses different training data, retrieval mechanisms, and synthesis logic. Testing one gives you less than half the picture.
• Real interface testing: Developer APIs return different results than the consumer-facing product. API responses often lack the web-search grounding that real sessions include. As we detail in our analysis of how AI visibility scores work, the measurement method matters as much as the measurement itself.

What you are looking for is your mention rate — the percentage of relevant buyer queries where your brand appears — broken out by platform, by query type, and over time. A brand at 40% mention rate that was at 55% three months ago is losing ground in ways a spot-check would never reveal.

How to shift the ratio

Once you know your actual inclusion rate and the specific factors suppressing it, you can target the gaps. The goal is not to game AI but to give it the evidence it needs to recommend you confidently.

• Cross the corroboration threshold. Identify the core claims AI needs to confirm about your brand — what you do, who you serve, why you are credible — and ensure at least 2–3 independent, authoritative sources confirm each one. This means earning editorial mentions, analyst coverage, or expert reviews that corroborate your positioning, not just mention your name.
• Eliminate positioning contradictions. Audit how your brand is described across every public source — your website, marketplaces, review sites, social profiles, press coverage. Where descriptions conflict, fix them. Where category labels differ, align them. AI’s confidence tracks directly with how consistently your sources agree about what you are.
• Close vocabulary gaps. Map the exact language buyers use when asking AI about your category and make sure your content uses those same terms. This is not keyword stuffing — it is ensuring AI can connect buyer intent to your brand without requiring a vocabulary translation it may not make.
• Unblock AI crawlers. Check your robots.txt for rules that block GPTBot, ClaudeBot, PerplexityBot, or Google-Extended. If you are blocking these, you are invisible to retrieval-augmented answers regardless of how strong your other signals are.
• Build across all three knowledge layers. Entity graph gaps require structured data, Wikipedia presence, and clear entity markup. Document graph gaps require fresh, authoritative web mentions. Concept graph gaps require consistent long-term presence in the types of content that feeds AI training data.
• Measure, then measure again. Run a baseline measurement of your inclusion rate. Make your changes. Wait 4–6 weeks for the effects to propagate. Measure again. The shift in your mention rate is the only reliable indicator of whether your efforts worked.

Last updated: April 2026