Do AI Engines Recommend the Same Companies? A 3-Engine Agreement Study

Why Engine Agreement Matters for AI Visibility

When a buyer asks an AI assistant for recommendations, they ask one assistant, not three. They might use ChatGPT because it is built into their workflow, Claude because their company has an enterprise contract, or Gemini because they are already in Google's ecosystem. The buyer does not run a consensus poll.

This creates a measurement problem for anyone tracking "AI visibility." If the three major engines recommend substantially different companies for the same question, then any score derived from a single engine describes that engine's opinion, not a stable property of the brand. A company might score well on one platform and poorly on another, and neither score would be wrong, just incomplete.

The stakes are practical. Marketing teams are beginning to treat AI recommendations the way they once treated search rankings: as a channel to monitor and optimize. Agencies sell AI visibility audits. Brands ask whether they are "showing up" in AI answers. But if the three engines disagree more often than they agree, then the answer to "are we showing up?" is always "it depends on where you look."

This study does not attempt to measure which engine gives better recommendations. That would require knowing the buyer's true preferences, which the data does not contain. Instead, it measures something narrower and checkable: when one engine names a company, how often do the others name it too?

Who this matters to

• Brands tracking AI visibility: A single-engine score may not generalize. A company visible on ChatGPT is not necessarily visible on Claude or Gemini.
• Agencies selling visibility services: Promising to improve "AI visibility" without specifying which engine is an underspecified promise.
• Buyers using AI for recommendations: The company you get depends on the assistant you ask. This is not a flaw in the assistants; it is a property of how recommendation engines work when they are trained on overlapping but not identical data.
• Researchers studying AI recommendations: Agreement rates provide a baseline for understanding how much engine choice matters versus question specifics.

How the Measurement Works

The unit of analysis is a (question, company, engine) triple. For each of the 2500 prompts in the sample, the study collected responses from all 3 engines. Every company named by at least one engine for a given question enters the denominator for that question.

The agreement calculation

For each company-question pair, the study counts how many engines named that company:

The denominator is 56819 pairs where at least one engine made a recommendation.

Design choices and their consequences

Deterministic sampling. Questions were selected by ordering prompt IDs by their MD5 hash, which produces a reproducible sample without storing a random seed. The same database will always yield the same sample.

Most recent run only. Some questions were run multiple times. Only the most recent completed run per (question, engine) was used, so a question re-run ten times contributes one observation, not ten.

Company universe per question. The denominator is not "all companies that exist" but "companies an engine was willing to recommend for this question." A company that no engine ever mentions for a given question does not appear in that question's denominator. This means the study measures disagreement among active recommendations, not the gap between recommendations and reality.

What this cannot see

The method cannot determine which engine is right. If ChatGPT recommends Company X and Claude does not, Company X might be a good recommendation Claude missed, or a bad recommendation ChatGPT should not have made. Agreement is not accuracy. High agreement could mean all three engines share the same bias.

The Headline Result: Engines Mostly Disagree

Of 56819 company-question pairs where at least one engine made a recommendation, 50209 (88.4 percent) were named by only one engine. Just 1812 pairs (3.2 percent) achieved unanimous agreement.

This is a statement about disagreement, not error. When ChatGPT recommends a company that Claude and Gemini do not mention, three explanations are possible:

1. ChatGPT found a good company the others missed. Perhaps its training data includes sources the others lack, or its retrieval system surfaced a relevant but obscure provider.

1. ChatGPT made a questionable recommendation. Perhaps it hallucinated a company name, or recommended a company that is no longer in business, or picked up on a marketing-heavy source that overstated the company's relevance.

1. The question has many valid answers and the engines are sampling from the space of good options. For a query like "best digital marketing agencies," there may be dozens of reasonable recommendations, and each engine picks a different subset.

The data cannot distinguish these cases. What it can establish is that relying on a single engine's recommendations, or measuring visibility on a single engine, captures less than half the picture.

The practical implication

If you measure your company's AI visibility on ChatGPT alone, you are measuring ChatGPT's opinion. That opinion may differ substantially from Claude's or Gemini's. 96 percent of companies that appear somewhere do not appear everywhere. This is not a rounding error; it is the dominant pattern.

A company that appears in all three engines for a question is rare enough to be notable. A company that appears in only one engine for a question is the norm.

The Shape of Disagreement: Binary, Not Gradual

The distribution of agreement is heavily skewed. Nearly nine in ten company-question pairs (88.4 percent) are sole mentions. Partial agreement, where exactly two engines name the same company, accounts for just 8.4 percent of pairs. Unanimous agreement is rarer still at 3.2 percent.

This distribution suggests that engines are not pulling from a shared list and occasionally disagreeing about edge cases. Instead, they appear to be constructing substantially different recommendation sets, with occasional overlap.

Engine-level differences

The engines do not recommend the same number of companies overall:

ChatGPT names the most companies, Gemini the fewest. The gap between them, roughly nine percentage points, is large enough that a company visible on the most expansive engine might be invisible on the most selective one.

Position when mentioned

When engines do mention companies, they tend to mention them at similar positions:

Gemini mentions companies slightly earlier on average (position 4.85), while Claude mentions them slightly later (position 5.38). The differences are small enough that position alone does not explain the disagreement. The engines are not agreeing on who to recommend and merely disagreeing on order; they are disagreeing on who to recommend at all.

Agreement Rates by Prompt Type

One might expect different levels of agreement for different kinds of questions. A navigational query ("How do I contact Acme Corp?") seems more constrained than a transactional query ("Who should I hire for mobile app development?"). Perhaps engines agree more on questions with fewer valid answers.

The data does not support this hypothesis. Appearance rates across prompt types are nearly identical:

The range spans just 38.6 percent to 36.9 percent, a gap of less than two percentage points. Navigational queries, which one might expect to have clearer "right answers," show the lowest appearance rate.

Why this might be

Several factors could explain the consistency:

1. All prompt types in this sample ask for company recommendations. Even a navigational query in this context involves naming a company. The engines face similar tasks regardless of prompt framing.

1. The sample may not include truly constrained queries. A query like "What is IBM's phone number?" has one right answer, but such queries may not appear in a sample designed to study company recommendations.

1. Engine disagreement may be driven more by training data differences than query characteristics. If the engines have access to different information about companies, that gap persists regardless of how the question is phrased.

The takeaway for practitioners: do not assume that clearer queries will produce more consistent recommendations across engines. The 88.4 percent sole-mention rate appears to be structural, not dependent on query type.

What Engines Are Reading When They Decide

The study does not have direct visibility into each engine's retrieval process, but the pattern of agreement suggests some hypotheses about what drives recommendations.

Shared information should produce shared outputs

If all three engines had access to identical information about every company, and used similar reasoning to evaluate that information, they would produce similar recommendations. The 88.4 percent sole-mention rate implies that at least one of these conditions fails.

Training data differences. Each engine is trained on a different corpus. A company with strong coverage in sources favored by ChatGPT's training may appear in ChatGPT's recommendations but not in Claude's or Gemini's. This is not a matter of one engine being better; it is a matter of different inputs producing different outputs.

Retrieval system differences. Modern AI assistants often combine base model knowledge with real-time retrieval. The retrieval systems may prioritize different sources or apply different relevance criteria.

Response generation differences. Even given identical retrieved information, the models may weight evidence differently. One might prioritize recency, another might prioritize source authority, a third might prioritize keyword matches.

What a company can influence

A company cannot control which sources each engine trusts. But it can increase its presence across the sources that engines collectively draw from:

• Authoritative directories that appear in multiple engines' training and retrieval pipelines
• Industry publications that engines treat as credible for the company's category
• Structured data (schema markup, knowledge graph entries) that helps engines verify company information
• Consistent entity information across the web, reducing the chance that an engine sees conflicting signals

The goal is not to optimize for one engine at the expense of others, but to build a presence that survives across different retrieval and ranking systems.

The Problem with Single-Engine Visibility Scores

Many AI visibility tools and audits measure presence on a single engine, often ChatGPT due to its market share. This study quantifies why that approach is incomplete.

The math of single-engine measurement

ChatGPT mentions 42.7 percent of the cohort. Claude mentions 39.9 percent. Gemini mentions 33.6 percent. But these are not subsets of each other. A company visible on ChatGPT is not necessarily visible on Claude.

29329 companies (96 percent of those visible anywhere) are visible on at least one engine but not all three. If you measure only on ChatGPT, you miss the companies that appear only on Claude or Gemini. More importantly, you may tell a client they are "visible" when they are visible on one engine and invisible on two.

What this means for visibility audits

A responsible AI visibility audit should:

1. Specify which engines were measured. "You are visible" should always be followed by "on ChatGPT, Claude, and/or Gemini."

1. Report per-engine results separately. An aggregate score hides whether a company has broad coverage or narrow.

1. Acknowledge what was not measured. Bing's Copilot, Perplexity, and other AI assistants may produce different results.

1. Update periodically. Engine behavior changes as models are updated and retrieval systems are refined.

What this means for brands

If a competitor is visible on all three engines and you are visible on one, you are at a disadvantage that a single-engine score would not reveal. Conversely, if you are invisible on the engine your best customers happen to use, your aggregate score may overstate your actual visibility to buyers.

The 88.4 percent sole-mention rate means that engine choice is not a minor detail. It is often the determining factor in whether a company appears at all.

What a Company on the Wrong Side Should Do

A company that appears on one engine but not the others faces a specific problem: whatever signals earned that single recommendation are not transferring to the other engines. This section outlines a diagnostic approach.

Step 1: Identify the gap

Before optimizing, understand the current state:

• Which engines mention your company, for which questions?
• Where in the response does your company appear?
• Are competitors mentioned that you are not?

This study provides data for 30564 companies across 2500 prompts. For companies in the dataset, the raw data shows which engines named them and when.

Step 2: Audit your information footprint

Engines draw from different sources. Audit your company's presence on:

• Major business directories (industry-specific and general)
• Review platforms relevant to your category
• Professional associations that maintain member directories
• Wikipedia and Wikidata (for larger companies)
• News coverage in outlets engines may cite

If you are strong in sources that ChatGPT favors but weak in sources Claude or Gemini favor, that gap may explain the disagreement.

Step 3: Fix entity consistency

Engines struggle when they see conflicting information. Common problems:

• Different company names on different platforms (Inc. vs. LLC vs. no suffix)
• Outdated addresses or phone numbers
• Inconsistent descriptions of services offered
• Multiple websites for the same company

Consistency helps engines verify that mentions of your company across sources refer to the same entity.

Step 4: Monitor, do not assume

Engine behavior changes. A company invisible today may appear after a model update. A company visible today may disappear. Build ongoing monitoring into your process, preferably across all major engines rather than the one you find most convenient.

The goal is not to game any single engine, but to be the kind of company that multiple systems, with different data and different reasoning, independently conclude is worth recommending.

How to Read the Published Dataset

The study covers 30564 companies across 2500 prompts, generating 170457 usable recommendation events. The data is structured to support both company-level and question-level analysis.

Key columns

• prompt_id: Unique identifier for the question. The same prompt_id across engines means the same question was asked.
• company: The company name as returned by the engine.
• engine: Which engine (ChatGPT, Claude, or Gemini) made this recommendation.
• position: Where in the response the company was mentioned (lower is earlier).
• prompt_type: The query classification (transactional, informational, commercial-investigation, or navigational).

Common analyses

Company-level visibility: Filter to a specific company and count how many prompts each engine recommended it for. Compare across engines to see where the company is strong or weak.

Question-level agreement: For a specific prompt_id, list all companies mentioned and by which engines. Calculate the agreement rate for that question.

Category analysis: Filter to prompts in a specific industry category. The data covers categories from Accessibility Testing Services (392 companies) to Web Development Services (506 companies).

Limitations of the dataset

• Snapshot in time. Data was collected between May and August 2026. Engine behavior may have changed since.
• No ground truth. The data shows what engines said, not whether they were right.
• Prompt sample, not exhaustive. 2500 prompts is a sample, not the universe of possible buyer questions.
• Company names as returned. The same company may appear with slightly different names across engines; normalization was applied but may not be perfect.

Reproducing the sample

The sample was selected by ordering prompt IDs by their MD5 hash. Any researcher with access to the same prompt database can reproduce the sample without needing a stored random seed.