How AI models decide which brands to recommend

When a buyer asks ChatGPT “What is the best project management tool for a growing team?”, the response is not a list pulled from a database. It is a narrative. The model constructs an argument: it defines what matters for that buyer, weighs trade-offs, and positions brands against criteria it has synthesised from its training data and any retrieved sources.

This matters because the narrative structure is where the decision happens. The model does not simply rank brands by popularity or mention frequency. It frames which criteria should matter, then evaluates brands against those criteria. A brand that appears in the response but is positioned against the wrong criteria has visibility without relevance.

The ontology challenge

We have spent the past year building a system that reads these narrative responses at scale: across models, across industries, and over time. The core challenge is not collecting the responses. It is making them comparable. Every model phrases things differently. The same criterion, “transparent pricing”, might appear as “clear cost structure”, “no hidden fees”, or “pricing visibility” depending on the model, the prompt, and the context.

This is where the ontology matters. We have built a categorisation layer that consolidates the natural language variation in AI responses into canonical criteria, elimination triggers, and hesitation signals. Across 18 projects spanning 13 industries, thousands of unique phrasings consolidate into 10 distinct decision criteria, 10 elimination trigger categories, and 8 hesitation signals.

The convergence is not coincidental. It reflects the fact that AI models, despite their different architectures and training data, arrive at remarkably similar decision frameworks when evaluating providers in a category.

The first thing an AI model does when answering a recommendation question is establish the decision frame. Before naming any brand, it tells the buyer what to look for.

This is not neutral. The criteria a model surfaces determine which brands can win the recommendation. If the model leads with “enterprise-grade security and compliance”, a startup with better UX but weaker security positioning is already at a disadvantage, regardless of how often it appears.

Three importance tiers

Across the industries we monitor, criteria cluster into three importance tiers. Of the 10 canonical criteria we track, roughly equal numbers are classified as high-importance and medium-importance by the models, with a small set classified as low-importance. The distribution is not static. Criteria shift importance over time, and they shift differently across models.

The strategic implication is direct. A brand’s AI visibility score tells you how often you appear. Criteria alignment tells you whether appearing helps or hurts. A brand positioned on medium-importance criteria while competitors own the high-importance ones is visible but structurally disadvantaged. The model is telling buyers to prioritise something the brand does not lead with.

What makes this actionable is that the criteria are observable and finite. There are not hundreds of them. Ten canonical criteria, monitored over time, give a brand a clear map of what the market is selecting for, and where their positioning is aligned or exposed. For a deeper look at how to track criteria alignment, see what to measure instead of AI visibility scores.

AI models do not just recommend. They eliminate. And the elimination logic is often more decisive than the recommendation logic.

When we track a brand’s journey through the buyer decision stages, a consistent pattern emerges. A brand can appear strongly in early exploration, where models list broad options, and then vanish in the elimination filter stage, where models apply hard constraints. The brand did not lose visibility. It was filtered out on a specific criterion.

Named, categorised, and trackable

We track 10 distinct elimination trigger categories across the industries we monitor. These are not vague concerns. They are specific, addressable reasons a model gives for excluding a brand: pricing transparency gaps, integration limitations, scalability questions, compliance gaps. Each trigger carries a severity rating (high, medium, or low) and a buyer reason that explains the logic.

The severity matters because high-severity triggers are binary. They do not reduce a brand’s ranking; they remove it from consideration entirely. A brand with strong positioning on every criterion except one high-severity elimination trigger is invisible to any buyer whose query activates that trigger.

This is the gap that visibility scores miss entirely. A brand can have high visibility in consideration-stage queries and zero presence in decision-stage queries if it carries an unaddressed elimination trigger. The aggregate score looks acceptable. The business impact is a complete loss at the point where decisions are made.

The fix is specific: identify which triggers apply to your brand, assess their severity, and address them with content that directly resolves the concern. The direct-answer content format is designed for exactly this: creating pages that address a specific buyer objection with verifiable evidence. The elimination filter is not a black box. The triggers are named, categorised, and trackable.

AI models need to justify their recommendations. A model that names a brand without supporting evidence is making an unsupported claim, and the architectures behind these models are increasingly designed to avoid that.

Trust in AI recommendations is not built the same way as trust in search rankings. It is not about backlinks or domain authority. It is about whether the model can construct a credible narrative around the brand. Does the brand have verifiable claims? Are third parties saying consistent things? Is there evidence beyond the brand’s own marketing?

A threshold, not a spectrum

What we observe across industries is that trust operates as a threshold, not a spectrum. A brand either has enough supporting evidence for the model to recommend it with confidence, or it does not. Brands below the threshold are mentioned cautiously, often with qualifiers (“some users report”, “it may be suitable for”) that signal low confidence to any buyer reading the response.

The trust signals that matter most are consistency and specificity. A brand that is described the same way across models, with specific claims about what it does well, has higher trust architecture than a brand whose description varies widely or relies on generic positioning. The models are synthesising from many sources. When those sources agree, the model’s confidence increases. When they disagree, the model hedges.

This has a practical consequence for content strategy. The goal is not to generate more content. It is to generate consistent, specific, verifiable content that multiple sources can echo. A single authoritative trust story that third parties reference does more for trust architecture than fifty blog posts that only the brand itself publishes.

The buyer’s decision journey through AI is not a funnel. It is a sequence of stages, each with its own logic, and brands can gain or lose ground at every transition.

We track six distinct decision stages: early exploration, deep evaluation, direct comparison, elimination filter, internal alignment, and final verification. At each stage, we measure two opposing forces: progression signals (evidence that the buyer is moving forward with a brand) and hesitation signals (evidence that the buyer is stalling or reconsidering).

The movement ratio

The ratio between progression and hesitation signals at each stage, what we call the movement ratio, is a direct indicator of friction. A movement ratio above 70% indicates smooth progression. Between 40% and 70%, there is moderate friction. Below 40%, the stage is a blockage point where the brand is actively losing ground.

Hesitation as a lead scoring signal

This is where the lead scoring potential becomes concrete. Consider a B2B SaaS company monitoring its AI visibility across buyer queries. The data shows strong progression through early exploration and deep evaluation (movement ratios above 70%), but the elimination filter stage shows 50% friction: equal numbers of progression and hesitation signals.

The hesitation signals at that stage are specific: “concerns about the long-term scalability of the solutions being considered”, “difficulty deciding between two providers that both meet most criteria but differ in one key aspect”, “uncertainty about whether the features of the less expensive provider are sufficient.”

Each hesitation signal is a lead scoring input. A prospect whose buyer journey matches this pattern is not lost. They are stuck at a known stage with a known concern. The content strategy response is equally specific: address the scalability concern with evidence, clarify the differentiating feature, reframe the price-value relationship. The competitor acknowledgment format works well here: addressing the comparison directly rather than ignoring it.

This turns AI visibility data into an intent signal. Not “this prospect visited our website” but “this prospect’s buyer journey is stalling at the elimination filter stage due to scalability concerns.” The first is a vanity metric. The second is actionable intelligence.

We track 8 distinct hesitation signal categories across the industries we monitor. The signals are not unique to individual brands. They recur across industries, which means the response patterns are reusable. A brand that addresses the three most common hesitation signals in its category has removed friction at the points where decisions stall most often.

One of the most common questions we hear is “which AI model matters most?” The data gives a clear answer: all of them, because they agree on more than they disagree on.

Across the industries we monitor, all three models surface the same 10 canonical decision criteria. Not “substantially similar” criteria. The same set: cost and fees, digital experience, expected outcomes, expertise and competence, flexibility and customisation, and five others. One hundred percent overlap.

Same criteria, different emphasis

Where they diverge is in how much weight each criterion receives. GPT is remarkably uniform: it assigns nearly equal signal density to every criterion, treating no single factor as more important than another. Claude is similarly balanced, with only marginal variation. Gemini is the outlier. It actively emphasises certain criteria over others: cost and fees receives over 50% more signal weight than independence and incentives.

Industry, product, and audience change the weighting

There is a critical caveat. While the criteria set is consistent at the category level, the weighting varies across industry, product type, and target audience. A cybersecurity provider and a recruitment platform both face “cost and fees” as a criterion, but the relative importance is different because the buyer’s risk calculus is different. The same applies within an industry: a product aimed at enterprise buyers triggers different emphasis patterns than one aimed at SMBs, even for the same service category. You cannot assume that the general pattern applies to your specific combination. You need to measure it.

The practical consequence is that a brand weak on cost transparency is more exposed in Gemini-influenced buyer journeys than in GPT-influenced ones. The criteria are the same, but the weighting creates different outcomes for different brands depending on where their positioning gaps sit.

This shapes the strategy. Optimising for one model’s specific weighting is fragile. When the model updates, the weighting shifts, and the positioning advantage disappears. The durable approach is to align with the full criteria set that all three models share, then monitor the model-specific emphasis patterns for early signals of shifting priorities. The shared criteria set is the foundation. The model-level variations are leading indicators of where the market is heading.

AI models are not search engines. They do not rank pages. They construct decision narratives that guide buyers through a structured evaluation process. Understanding that process, the criteria weighting, the elimination filters, the trust thresholds, the stage progression, the cross-model patterns, is what separates brands that appear from brands that are chosen.

The companies gaining ground in AI-driven discovery are not the ones with the highest visibility scores. They are the ones that understand the decision logic, monitor how it shifts, and align their positioning with what the market is actually selecting for.

If you are measuring AI visibility today, ask yourself: do you know which criteria the models are telling buyers to prioritise for your specific industry, product, and audience? Do you know which elimination triggers apply to your brand? Do you know at which decision stage your buyers hesitate, and why?

These answers vary by industry, by service type, and by buyer segment. The patterns described in this article hold across the 13 industries we monitor, but the specific weighting, the severity of individual triggers, and the friction points in the journey are different for every combination. General benchmarks are not enough. You need continuous measurement for your market, your positioning, and your competitive set.

If you are not measuring those things, you are optimising for the wrong metric. For the five metrics that do predict business impact, see what to measure instead of AI visibility scores.