We Asked Two AI Models About 24 Brands. Their Answers Did Not Match.

Your brand has an AI reputation. It changes with the model your customer uses, the day they ask, and the way they phrase the question. On 20 February 2026 we asked two AI models and three AI search tools about 24 well-known companies in 8 industries. This article reports what those answers showed.

Revised on 4 November 2026. This version corrects how the study was labelled and how its source numbers were described. The correction note at the end lists every change.

---

Why This Study Matters

Every time a buyer asks ChatGPT, Gemini or Perplexity about your industry, the tool writes a description of your brand on the spot. That description is your AI reputation. You cannot read it the way you read press coverage, and the next answer can differ from the last one.

We set out to answer four questions. Which web pages do AI search tools cite when they describe a brand? How consistent are the answers of a plain AI model? Does the industry change the picture? Do listed companies get a different treatment from private ones?

---

What We Tested

We studied 24 well-known companies across 8 industries: E-commerce, Healthcare Technology, Logistics, Consulting, HR Tech, Web Hosting, Marketing Technology and FinTech. Three companies per industry; 12 listed and 12 private. Rankfor.AI is one of the 24.

The study has two arms. Both were collected on 20 February 2026.

Arm 1: answers without web search. GPT-5.2 and Gemini 3 Flash answered 15 questions about each company from their own memory, with no web search. Three of the questions were asked five times each to measure consistency. 1,296 calls were planned. The archive holds 1,311 answers (663 from GPT-5.2, 648 from Gemini 3 Flash), because 15 calls for two companies were recorded twice, and every row stayed in the analysis. These answers cite no sources.

Arm 2: answers with web search and cited links. Three AI search tools, Gemini 3 Flash with Google Search, Grok 3 with web search and Perplexity Sonar Pro, answered five questions about each company and were asked to list the sources they drew from. They returned 3,041 links in total: 1,311 from Gemini, 1,019 from Grok and 711 from Perplexity. GPT-5.2 took no part in this arm.

How the links were sorted. A fixed list of domain patterns placed each link into one of nine categories: company-owned, tier-1 news, review platform, trade or sector news, Wikipedia, financial or regulatory, social media, LinkedIn, government. A link whose domain matched no pattern was labelled "other". No person coded the links by hand.

How the answers were scored. Sentiment is a VADER lexicon score from -1 to +1: it counts positive and negative words in the answer. Consistency is word-overlap (TF-IDF) similarity between the five answers to the same question.

---

Finding 1: Where the Cited Links Pointed

Of the 3,041 links cited by the three search tools, 14.7% pointed to the company's own domain, 5.7% to tier-1 news and 4.6% to review platforms. 70.2% pointed to domains outside the study's named categories.

Where the link pointed	Share of 3,041 links
Company-owned domain	14.7%
Tier-1 news media	5.7%
Review platforms	4.6%
Trade or sector news	1.8%
Wikipedia	1.5%
Financial or regulatory	0.6%
Social media	0.5%
LinkedIn	0.4%
Government	0.1%
Outside the named categories ("other")	70.2%

The 70.2% needs two sentences of explanation. 1,311 of those 2,134 links, 43.1% of all links, are the Gemini rows, and they carried no usable domain: the tool returned a redirect address (1,195 links) or a string of page markup (116). The other 823 came from Grok and Perplexity and had readable domains the list did not cover, such as youtube.com (36), techradar.com (33), reddit.com (30), statista.com (28) and bbc.com (22).

Among the 907 links that matched a named category, 49.3% pointed to the company's own domain.

What this means for you: When an AI search tool described one of these 24 brands and cited a page, the company's own website was the largest single named destination. Tier-1 news was second. The 70.2% describes how short the category list was and how Gemini reported its links. It says nothing about where a model's memory comes from.

---

Finding 2: The Two Models Scored the Same Brands Differently

We compared how GPT-5.2 and Gemini 3 Flash described the same 24 companies in Arm 1.

GPT-5.2 had a mean sentiment score of 0.83. 97.4% of its answers scored as positive.

Gemini 3 Flash had a mean sentiment score of 0.32. 51.1% of its answers scored as positive.

The gap is 0.51 points. The correlation between the two models' company-level scores was r = 0.37 (p = 0.07, 24 companies), which is weak and not statistically significant. Sentiment here is a VADER lexicon score: it counts positive and negative words in the answer. No person judged the brands.

What this means for you: If you check your brand on ChatGPT alone, you see one model's tone. In this sample the other model was markedly cooler on the same companies.

---

Finding 3: Ask the Same Model Twice and You Get Different Answers

We asked three questions five times each for every company on both models, 144 question groups in total. Mean word-overlap similarity between the five answers in a group was 0.54.

Only 2.8% of the groups (4 of 144) reached the 0.85 level we defined as "stable". 93.8% (135 of 144) fell below 0.70, which we classified as "variable".

GPT-5.2 was slightly more consistent (0.57) than Gemini 3 Flash (0.51). E-commerce had the lowest average (0.49), Consulting the highest (0.57).

What this means for you: A single answer is one draw. In this sample, five draws of the same question rarely used the same words. Repeat the question before you read anything into one answer.

---

Finding 4: The Industry Changed Which Links Were Cited

The industry of the company and the category of the cited link were related (chi-square = 143.28, 63 degrees of freedom, p < 0.001). The effect is small (Cramer's V = 0.08), and two caveats apply: several table cells have expected counts below 5, and links from the same answer are not independent observations. Read the sector figures as a description of this sample.

Consulting had the highest share of company-owned plus tier-1 news links: 30.4% (108 of 355). McKinsey had 19 tier-1 news links out of its 154 cited links (12.3%), or 35% of its 54 links in a named category.

Marketing Technology had the highest share of links outside the named categories: 81.4% (210 of 258).

HR Tech had the highest share of review-platform links: 9.5% of all its links (39 of 410), or 27.3% of its links in a named category. G2 and Capterra were among those domains.

What this means for you: Which pages the search tools cited differed by sector. A consulting firm's cited pages leaned on its own site and the business press. An HR Tech vendor's cited pages included review platforms. Check the pattern for your own sector before you copy a generic playbook.

---

Finding 5: Listed and Private Companies Looked the Same

We compared the 12 listed companies with the 12 private ones on every measure in the study. No difference reached significance.

• Consistency (Arm 1): listed 0.54, private 0.54 (d = 0.03)
• Positive answers (Arm 1): listed 74.4%, private 74.6%
• Mean sentiment (Arm 1): listed 0.57, private 0.57 (d = 0.01)
• Spread of link categories (Arm 2): Shannon entropy 1.43 listed, 1.37 private (p = 0.59)
• Cited links per company (Arm 2): 123.4 listed, 130.0 private (p = 0.46)
• Financial or regulatory links (Arm 2): 0.7% listed, 0.6% private

Stock-exchange filings, Bloomberg coverage and Crunchbase profiles did not give the listed companies a measurable edge in this sample.

What this means for you: All 24 companies are established names. With 12 per group, the test can detect large differences only. Within that limit, being listed changed nothing we measured.

---

The Three Search Tools Compared

Each search tool reported its links differently, and the share of links outside the named categories differed by tool:

• Gemini 3 Flash with Google Search: 100% (1,311 of 1,311). Every link came back as a redirect address or page markup, so no domain could be read.
• Perplexity Sonar Pro: 60.6% (431 of 711).
• Grok 3: 38.5% (392 of 1,019).

An earlier version of this article said that search mode shrinks the unattributable share dramatically. The file shows Perplexity's own share outside the named categories at 60.6%. What a search tool gives you is a list of pages you can open and read. It does not give you a cleaner category split.

---

What a Marketing Team Does With This

Four steps, in order:

1. Compare the answer with brand facts. Ask each model and search tool the questions your buyers ask. Note where the answer is wrong, outdated or missing something you would expect a buyer to hear.
2. Inspect which sources the tools cite. Open every cited page. Record which are yours, which are third-party, and what each one says about you.
3. Agree which information can be improved. Some cited pages are yours to edit. Others belong to reviewers, the press or a directory, and you can ask, correct or add to them.
4. Repeat the measurement. Ask the same questions again after the changes, several times each, and compare.

One caution. A cited source does not prove that it caused the whole answer, and changing it does not guarantee a recommendation. In this sample, five repeats of one question rarely matched, so compare runs, never single answers.

Advice from us, outside what the study measured: pick a fixed interval for the repeat and keep it.

These four steps are the workflow of our managed AI search service, where a Rankfor.AI team runs the measurement and the source review for you.

---

Study Details

• Cohort: 24 well-known companies, 8 industries, 3 per industry; 12 listed, 12 private
• Collection date: 20 February 2026, both arms
• Arm 1, answers without web search: GPT-5.2 (663 answers) and Gemini 3 Flash (648 answers); 15 questions per company; 3 questions repeated 5 times; 1,311 answers in total
• Arm 2, answers with web search: Gemini 3 Flash with Google Search, Grok 3 with web search, Perplexity Sonar Pro; 5 questions per company per tool; 3,041 cited links
• Link categories: nine named categories from a fixed domain list plus "other"; no hand coding
• Sentiment: VADER lexicon compound score; positive at 0.05 and above
• Consistency: word-overlap (TF-IDF) similarity, mean over the pairs in each group of five answers
• Publication status: unpublished Rankfor.AI study; not peer reviewed; no journal version
• Data: the answers and link records are held in the Rankfor.AI research archive and have not been deposited as a public dataset

---

Correction Note, 4 November 2026

This article was first published on 21 February 2026 and revised on 4 November 2026. The revision is an editorial correction of the original text against the study's own archive. No new data were collected.

What changed:

• Publication status. The earlier text described the study as peer reviewed and published in a journal. It is an unpublished Rankfor.AI study with no journal version. The label has been removed everywhere on this page.
• What the 3,041 records are. The earlier text called them "source attributions" coded by hand by two coders. They are links cited by three AI search tools, sorted by a fixed domain list. No hand coding took place.
• What 70.2% means. The earlier text said that 70.2% of AI brand knowledge cannot be traced to any source. The figure is the share of cited links whose domain matched none of the study's named categories, and 43.1% of all links are Gemini rows with no readable domain. The claim about untraceable training data has been withdrawn.
• Search mode. The earlier text said that search mode shrinks the unattributable share dramatically. Perplexity's own share outside the named categories is 60.6%.
• Numbers. 96.4% is now 97.4% (GPT-5.2 positive answers); 51.5% is now 51.1% (Gemini); 73.9% is now 74.4% (listed, positive); 30.8% is now 30.4% (Consulting); 81.3% is now 81.4% (Marketing Technology); 9.6% is now 9.5% (HR Tech review platforms); "72 grounded probes via Perplexity" is now 360 search-tool answers across three tools; the McKinsey sentence now reports its own counts.
• Method labels. "Cosine similarity" is now word-overlap (TF-IDF) similarity; sentiment is labelled as a VADER lexicon score; the chi-square caveats on small expected counts and non-independent links have been restored.
• Scope statements. "Largest multi-industry study to date", "AI treats a startup the same as a Fortune 500 company" and "the playing field is level" have been removed. The cohort is 24 well-known companies.
• Title. The earlier title said "3 AI platforms". The comparison in Finding 2 is between two models; the search arm used three tools.

The wording published between 21 February and 4 November 2026 is kept in the Rankfor.AI editorial archive, and the Wayback Machine holds captures of the earlier page from 15 June 2026 onward.

A later study with a different cohort and models, "Measuring Corporate Reputation in the Age of AI" (24 companies, March 2026, GPT-5.4 and Gemini 3.1 Pro Preview), has an open replication package at doi.org/10.5281/zenodo.21796757. It is a separate study. Its numbers are not the numbers on this page.

---

The Bottom Line

On one day in February 2026, two AI models described the same 24 companies in different tones, five answers to the same question rarely matched, and three AI search tools cited the company's own website more often than any other named category. The brands that win AI recommendation share will be the ones that measure this repeatedly and work on the pages the tools cite.