---
title: "“Is this brand legit?” How AI assistants build a reputation"
description: "Asked “Is this brand legit?”, four AI engines said yes in every complete answer, and 99.7% raised a problem. How they build a reputation from evidence."
canonical: "https://underneath.agency/research/is-it-legit-ai-reputation-study"
published: 2026-09-26
updated: 2026-10-08
publisher: "Underneath (https://underneath.agency/agent)"
entity: "https://underneath.agency/.well-known/entity.json"
---
Research · AI assistants

# “Is this brand legit?” How AI assistants build a reputation

Before buying, people ask AI assistants whether a company can be trusted. We asked ChatGPT, Gemini, Perplexity and Google AI Mode “Is {brand} legit? What do customers say about it?” for 79 brands across eight industries on 26 September 2026. This version of the study looks past which websites the answers cite. It asks how the engines construct a brand’s reputation from that evidence: which kinds of source they draw on, what positive and negative claims they make, where the negative claims appear, and whether the cited sources support those claims.

Every complete answer said the brand was legitimate, and almost every one then set out problems. Legitimacy and reputation behave as two separate questions: the first always gets a yes, the second never gets an unqualified one. The engines build that reputation from different evidence, and the differences hold after accounting for the brands being asked about.

## The short version

1. **Legitimate, but.** All 312 complete answers affirmed that the brand was legitimate, according to both coders, and 99.7% of them made at least one negative claim about it. No answer painted a purely positive picture: 61.2% gave substantial concerns (95% interval 53.7% to 68.2%) and 38.8% mild ones.
2. **The problems are there, but rarely first.** Of the 5,188 reputation claims in the answers, 35.5% were negative. Negative claims were in the first paragraph of 10.9% of answers: 31.6% for Perplexity, 1.3% for Gemini and for Google AI Mode.
3. **Review platforms carry the negative claims.** 88.0% of answers cited a review or complaint platform. Claims attached only to review platforms were negative 56.5% of the time. Claims attached only to editorial review sites were negative 19.9% of the time, and those attached only to the brand’s own website 6.4%.
4. **The review evidence is concentrated.** Trustpilot and the BBB account for 61.7% of all review-platform citations. Across all sources, 484 domains were cited but the effective number of sources is 30.5.
5. **Each engine has its own evidence regime.** After allowing for brand, brand group and industry, Gemini was far less likely than ChatGPT to cite the BBB (odds ratio 0.05) and Google AI Mode far more likely to cite Reddit (5.76). Perplexity cited the brand’s own site in 94.9% of answers, Google AI Mode in 17.7%.
6. **Citations mostly support the claims they are attached to, when they can be checked.** In a random sample of 240 cited claims, a readable cited page supported the claim fully or in part 72.4% of the time. But 71.4% of claims that said a problem was common rested on evidence that did not show how common it was. Trustpilot, the BBB and ConsumerAffairs block automated reading, so 119 of the 240 claims could not be checked.

## Research questions

| Question | Answered here? |
|---|---|
| RQ1. Which classes of source do the engines cite when judging a brand, and how concentrated are they? | Yes |
| RQ2. Do the engines draw on different evidence, net of brand, brand group and industry? | Yes |
| RQ3. Which positive and negative claims do answers make, and how prominent are the negative ones? | Yes |
| RQ4. Do the cited sources support the claims attached to them? | For a sample, with bounds |
| RQ5. Is the reputation an engine gives a brand stable across runs, wordings and dates? | No: one run per engine and brand |
| RQ6. Are source classes associated with the kind of claims made? | Descriptively only |

## Five parts of an AI reputation

We treat an answer to “Is this brand legit?” as a reputation with five parts, coded separately for every answer, rather than as a single verdict:

- **Legitimacy:** whether the answer says the business is real and legitimate (not whether it is good), and whether its first sentence says so.
- **Overall structure:** positive; legitimate with mild concerns; legitimate with substantial concerns; ambiguous; negative; or undetermined.
- **Complaint profile:** the negative claims, by topic (customer service, billing and fees, product quality, legal and regulatory, ratings), and where in the answer they appear.
- **Evidence:** the sources cited, by class, and which claims each source is attached to.
- **Uncertainty:** whether the answer qualifies the evidence itself, for example by noting a small review sample or that experiences vary by franchise.

This follows the argument that visibility in generative search is a set of distinct measures, not one number. Here the measures are what an engine says about a brand and on what evidence.

## What we analyzed

The data are the 316 answers collected on 26 September 2026 (79 brands, four engines, one run each). No new questions were asked for this version.

- **Brands:** in each of eight industries, five brands the assistants named most widely in our [four-assistant study](https://underneath.agency/research/ai-assistants-brand-agreement-study) and five named by only one assistant (four in home and local services). 21 names were completed or corrected by hand.
- **Claims:** every answer was split into atomic reputation claims, 5,188 in all, each with its polarity, topic, the sources attached to it, its paragraph and how it described prevalence. Each answer was also coded on the five parts above. The coder was Claude Opus, reading the answer with its citations replaced by neutral tags and without being told the engine.
- **Second coder:** Claude Sonnet independently coded the answer-level measures for all 316 answers, and the polarity and topic of the 1,382 claims in 80 randomly drawn answers.
- **Sources:** each of the 484 cited domains was assigned one of eight classes. Domains belonging to the brand or its parent company were identified brand by brand and counted as the brand’s own website.
- **Support:** 60 cited claims per engine were drawn at random. Each was judged against the cited page, fetched on 28 September, or against the passage that Gemini’s link quoted. Two models judged each claim independently.

All coding and judging on this page was done by AI models. No person coded the sample. Answers that were incomplete (4, all ChatGPT) or cited nothing (3 ChatGPT, 3 Gemini) stay in the data as outcomes.

## Study 1: the evidence the engines draw on

### Source classes

| Share of answers citing at least one | All (95% interval) | ChatGPT | Gemini | Perplexity | Google AI Mode |
|---|---|---|---|---|---|
| Review or complaint platform | 88.0% (83.9% to 91.5%) | 92.4% | 68.4% | 97.5% | 93.7% |
| The brand’s own website | 51.6% (47.2% to 56.3%) | 70.9% | 22.8% | 94.9% | 17.7% |
| Editorial review or comparison site | 40.8% (34.2% to 47.2%) | 7.6% | 48.1% | 74.7% | 32.9% |
| Another company’s website | 37.7% (32.0% to 43.4%) | 20.3% | 44.3% | 69.6% | 16.5% |
| Forum, social or video | 28.2% (22.5% to 34.5%) | 11.4% | 26.6% | 31.6% | 43.0% |
| News media | 17.7% (13.9% to 21.5%) | 1.3% | 6.3% | 54.4% | 8.9% |
| Encyclopedia or scholarly reference | 13.9% (10.1% to 17.7%) | 0.0% | 6.3% | 43.0% | 6.3% |
| Government or regulator | 8.2% (5.4% to 11.7%) | 15.2% | 0.0% | 15.2% | 2.5% |
| Consumer advocacy nonprofit | 8.2% (4.4% to 12.7%) | 5.1% | 3.8% | 16.5% | 7.6% |

The individual sites behind these classes are unchanged from the first version: trustpilot.com was cited in 65.8% of answers, bbb.org in 51.3%, reddit.com in 24.7% and consumeraffairs.com in 22.5%.

Counted by citations rather than answers, review platforms made up 56.3% of ChatGPT’s citations and 53.3% of Google AI Mode’s, against 24.3% of Perplexity’s. Perplexity spreads its citations across more classes: a median of 5 source classes per answer, against 2 for each of the other engines.

### Concentration

Counting each domain once per answer, the 316 answers made 1,667 citations to 484 domains. The Herfindahl-Hirschman index, which sums squared shares, is 0.0328, so the answers behave as if they drew evenly on 30.5 sources. The review evidence is far more concentrated. 36 review platforms were cited, but Trustpilot and the BBB account for 61.7% of review-platform citations, and the effective number of review platforms is 4.7.

| Concentration of cited domains | ChatGPT | Gemini | Perplexity | Google AI Mode |
|---|---|---|---|---|
| Distinct domains | 104 | 134 | 376 | 67 |
| Share of the top 3 domains | 48.5% | 21.3% | 18.3% | 48.6% |
| Effective number of sources | 10.4 | 35.4 | 54.9 | 10.6 |

### More citations do not mean more independent evidence

The number of domains an answer cites and the number of source classes it draws on move together (rank correlation 0.84). They are not the same thing. 20.3% of ChatGPT answers drew on a single class of source, and several sites can repeat the same customer reviews. We did not measure whether cited sources are independent of one another, so source classes are only a proxy for diversity of evidence.

### Engine differences, net of the brands asked about

Answers about the same brand are not independent, so we modeled each outcome with brand clusters, adjusting for brand group and industry (odds ratios against ChatGPT, 95% intervals).

| Outcome | Gemini | Perplexity | Google AI Mode |
|---|---|---|---|
| Cites Trustpilot | 0.21 (0.1 to 0.44) | 2.72 (1.25 to 5.91) | 1.5 (0.79 to 2.85) |
| Cites the BBB | 0.05 (0.02 to 0.11) | 1.0 (0.61 to 1.65) | 0.19 (0.1 to 0.36) |
| Cites Reddit | 1.87 (0.79 to 4.42) | 3.15 (1.34 to 7.42) | 5.76 (2.58 to 12.88) |
| Cites the brand’s own website | 0.11 (0.06 to 0.24) | 8.09 (2.72 to 24.11) | 0.08 (0.04 to 0.17) |
| Substantial concerns | 0.21 (0.11 to 0.4) | 1.06 (0.57 to 1.98) | 0.23 (0.12 to 0.43) |
| Negative claim in the first paragraph | 0.13 (0.02 to 0.94) | 5.28 (2.11 to 13.23) | 0.13 (0.02 to 0.94) |
| Qualifies its own evidence | 0.22 (0.1 to 0.47) | 0.28 (0.14 to 0.56) | 0.1 (0.04 to 0.21) |

No Gemini answer cited a government or regulator source, so that outcome is not modeled. A model with a random intercept for each brand gave the same direction for every engine effect. Brand group had no clear effect on any of these outcomes, and no engine-by-group interaction was significant. We read the engines as different evidence-selection regimes, not as better or worse.

## Study 2: what the answers say

### The shape of the answers

| Share of complete answers | All | ChatGPT | Gemini | Perplexity | Google AI Mode |
|---|---|---|---|---|---|
| Legitimate, substantial concerns | 61.2% | 76.0% | 45.6% | 77.2% | 46.8% |
| Legitimate, mild concerns | 38.8% | 24.0% | 54.4% | 22.8% | 53.2% |
| Positive, ambiguous, negative or undetermined | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| First sentence affirms legitimacy, with or without a qualifier (all answers) | 97.2% | 88.6% | 100.0% | 100.0% | 100.0% |
| Qualifies its own evidence | 66.0% | 88.0% | 63.3% | 68.4% | 45.6% |

Perplexity’s first sentence carried a qualifier (“legitimate, but reviews are mixed”) in 31.6% of answers. The ChatGPT answers without a first-sentence verdict (11.4%) opened with a preamble or a list of branches; none disputed that the company was real.

### Legitimate, but

Among the 312 answers that affirmed legitimacy, the median answer gave 33.3% of its claims to negatives. In 20.2% of answers (95% interval 14.6% to 25.6%) the negative claims outnumbered the positive ones. An answer can open “Yes, this is a legitimate company” and spend most of what follows on complaints. A yes to “Is it legit?” says little about the reputation the answer then describes.

### What the claims are about

| Claim topic | Share of all claims | Share of these claims that are negative |
|---|---|---|
| Product or service quality | 22.0% | 37.2% |
| General customer experience | 16.1% | 27.2% |
| Identity and track record | 12.9% | 4.5% |
| Billing, fees, refunds and cancellation | 12.5% | 71.2% |
| Customer service | 12.0% | 69.7% |
| Legitimacy | 11.4% | 0.3% |
| Ratings on named platforms | 8.0% | 36.5% |
| Legal, regulatory or formal complaint record | 2.1% | 81.1% |
| Expert or press assessment | 1.9% | 8.0% |

| Share of complete answers with a negative claim about | All | ChatGPT | Gemini | Perplexity | Google AI Mode |
|---|---|---|---|---|---|
| Customer service | 76.6% | 81.3% | 65.8% | 84.8% | 74.7% |
| Billing, fees or cancellation | 70.2% | 72.0% | 70.9% | 64.6% | 73.4% |
| Product or service quality | 66.3% | 72.0% | 67.1% | 62.0% | 64.6% |
| Legal, regulatory or formal complaint record | 20.8% | 40.0% | 2.5% | 40.5% | 1.3% |

Negative claims were mostly stated as general patterns. 66.0% said a problem was common or frequent without a number, 11.4% gave a number or rating, and 14.1% described individual cases.

### Which sources carry which claims

| Claims whose citations are all from one class | Claims | Negative | Positive |
|---|---|---|---|
| Review or complaint platform | 1,542 | 56.5% | 33.9% |
| Forum, social or video | 140 | 40.0% | 45.7% |
| Another company’s website | 210 | 30.5% | 59.0% |
| Editorial review or comparison site | 351 | 19.9% | 65.8% |
| The brand’s own website | 377 | 6.4% | 66.3% |
| No citation attached | 1,507 | 24.4% | 54.7% |

The engines attached a citation to 71.0% of claims: 93.3% for Perplexity, 56.7% for Google AI Mode. Review platforms supply most of the negative claims that carry a citation. Brand-owned pages supply claims about identity and legitimacy. This is an association between source class and claim type in the answers. It does not show that a source caused a claim.

### Do the cited sources support the claims?

| 240 sampled cited claims (60 per engine) | Claims | Supported or partly supported |
|---|---|---|
| No cited source could be read (mostly Trustpilot, BBB, ConsumerAffairs) | 119 | not judged |
| Readable, but not about the claim’s company or empty | 8 | not judged |
| Judged against fetched cited pages | 58 | 72.4% (62.3% to 82.5%) |
| Judged against fetched pages, every cited source readable | 25 | 84.0% |
| Judged against Gemini’s quoted passages only | 55 | 45.5% (33.3% to 57.1%) |

Against fetched pages, 46.6% of claims were fully supported, 25.9% partly supported and 25.9% not supported; 1.7% were contradicted. When a claim cited several sources and some were blocked, the judge saw only the readable ones, so “not supported” is an upper bound. That is why the rate rises to 84.0% when every cited source could be read. Gemini’s links quote only the start and end of a passage, so its lower figure is partly a limit of the evidence. Across all 240 claims, the supported share lies between 27.9% and 80.8%, depending on how the claims that could not be checked would come out.

The clearest gap concerns prevalence. Of the 63 checkable claims that described a problem as common, 71.4% cited evidence that showed individual reports, or nothing about frequency. The two judges agreed on 81.8% of verdicts (kappa 0.68, supported-or-partly against not).

### Widely named and rarely named brands

Whether a brand was widely named in our earlier study made little difference to how its reputation was built. The difference between the two groups was within the margin of error for citing a review platform, citing the brand’s own site, giving substantial concerns, and putting a negative claim first. The exception was billing: 80.6% of answers about widely named brands made a negative billing claim, against 59.2% for brands named by one assistant (difference 95% interval 7.6 to 35.5 points). This is an association in 79 brands, not an effect of being well known.

### The same brand, four engines

For the 75 brands with four complete answers, all four engines gave the same overall structure for 38.7%. Two engines answering about the same brand shared on average 0.2 of their cited domains (Jaccard index). Different engines give the same brand noticeably different evidence and a different weight of concerns.

## Observed, inferred and unknown

**What we observe.** Every complete answer says the brand is legitimate. Almost every one then makes negative claims, mostly about customer service, billing and product quality, and mostly as general patterns. Review platforms, above all Trustpilot and the BBB, are cited in most answers and carry most of the cited negative claims. The engines differ systematically in which classes of source they cite and in how prominently they place concerns.

**What we infer.** The engines treat “is it legitimate” and “is it good” as separate questions: the first is answered from identity evidence, often the brand’s own pages, and the second from customer-review evidence. The reputation a buyer receives therefore depends on the engine as well as on the brand.

**What remains unknown.** Whether the same engine would give the same reputation on another run, with other wording or on another day. Whether a source that is cited changed the answer. Whether the negative claims reflect the balance of all customer experience or only what review platforms collect. Whether several cited sites repeat the same underlying reviews.

## What this means

The points below are our interpretation. They follow from the findings but were not tested.

- **Being called legitimate is not the finish line.** Every brand in the sample passed that test. What differed was the complaint profile that followed, and that profile draws most heavily on review platforms.
- **Recurring themes on review platforms are the likeliest content of an AI reputation.** Billing, cancellation and support problems appeared in most answers. Fixing them where customers report them addresses the evidence the engines cite.
- **Official pages supply the identity half.** Claims citing the brand’s own site were overwhelmingly about legitimacy and track record. A clear page on who the company is and how it handles complaints gives engines accurate material for that part of the answer, especially for Perplexity and ChatGPT, which cite brand sites most often.
- **Check the engine, not just “AI”.** A brand can get a mostly positive answer from Gemini and a concern-heavy one from Perplexity on the same day.

## Where this sits in GEO research

Research on generative engine optimization began with visibility: whether changing content changes how often it appears in AI answers. Later work compared which sources different AI search systems draw on, and argued that visibility should be measured as several separate quantities (discoverability, citation, claim support, stability) rather than one rank. This study applies that view to a consequential task, judging whether a business can be trusted. It measures source selection, claim construction and claim support in live answers. It does not measure stability over time or causal influence. Those need repeated runs and controlled experiments that add or remove a source.

## Methodology

- **Question and engines:** “Is {brand} legit? What do customers say about it?”; ChatGPT and Gemini consumer apps (DataForSEO LLM Scraper), Google AI Mode (DataForSEO SERP API), Perplexity sonar (DataForSEO LLM Responses API); US location; one run each; 26 September 2026.
- **Brands:** 79 from the national questions of the four-assistant study. Per industry: the five named by the most assistants, and five drawn at random (seed 20260926) from brands named by one assistant.
- **Claim and answer coding:** Claude Opus coded each answer with citations replaced by neutral tags and the engine hidden. It recorded status, first-sentence frame, legitimacy, structure (six categories), uncertainty, and every atomic reputation claim with its polarity, topic, attached sources, paragraph and prevalence wording.
- **Second coder:** Claude Sonnet, blind to the engine and to the first coder’s answer-level codes. Agreement: overall structure 84.2% (kappa 0.68), first-sentence frame 89.9% (kappa 0.63), negative claim in the first paragraph 92.0% (kappa 0.64), uncertainty 72.5% (kappa 0.48), and legitimacy affirmed or not 100.0%. On 1,382 claims, polarity agreed 93.2% (kappa 0.89) and topic 82.8% (kappa 0.8). The uncertainty measure is the least reliable and should be read as indicative.
- **Sources:** registrable domains, each coded into one of eight classes by Claude Opus (the list is in the dataset). Domains owned by the brand or its parent, identified per brand, count as the brand’s own website. The first version matched the brand’s name inside the domain, which found own sites in 39.9% of answers; recognizing parent companies and differently named official sites raises this to 51.6%.
- **Concentration:** each domain counted once per answer; Herfindahl-Hirschman index over domain shares; effective number of sources = 1 / index.
- **Support:** 60 cited claims per engine drawn at random (seed 20260928). Evidence was each cited page fetched over plain HTTP on 28 September 2026, or the passage that Gemini’s link quoted. Claude Opus judged each claim, and Claude Sonnet judged it independently. Claims with no readable source stay in the denominator, and bounds are given. Of the 2,414 cited pages, 38.6% could be read, including 6.6% of review-platform pages.
- **Models:** GEE logistic regression with brand clusters, fixed effects for engine, brand group and industry, and cluster-robust intervals. Check: a logistic model with a random intercept for each brand.
- **Intervals:** 95% bootstrap resampling brands (2,000 resamples).
- **Update schedule:** quarterly.

## Limitations

- One answer per engine and brand, on one date. How stable an engine’s account of a brand is across runs, wordings and days is not measured, and single answers vary.
- Citation is not influence. The data show which sources the engines cite and which claims they attach to them, not that those sources caused the answer.
- All coding and judging was done by two AI models from the same family. Their agreement is a check on consistency, not on accuracy. No person coded the sample.
- The review platforms cited most often block automated reading, so claim support could mostly be checked for other sources. Pages were fetched two days after the answers.
- Whether several cited sources repeat the same underlying reviews was not measured.
- Perplexity was tested through its API. The brand sample leans toward well-known US companies.

## What changed in version 1.1

Version 1.0 (26 September 2026) counted which websites were cited and flagged five kinds of warning with phrase rules. Following an external review, version 1.1 re-analyzes the same 316 answers. It adds claim-level coding, the five-part framework, a source taxonomy with concentration measures, a claim-support check, models that account for answers about the same brand, and a second coder.

The source figures are unchanged. The phrase-rule figures are replaced. The rules caught narrow wordings: they found billing problems in 27.8% of answers and customer service complaints in 30.1%, where claim coding finds negative billing claims in 70.2% and service claims in 76.6%. The complaint rule itself agreed with claim coding on 95.5% of answers. The earlier statement that Trustpilot and BBB profiles shape the answer, and that “what those sites show is what AI repeats”, is withdrawn. The data show that those platforms are a prominent part of the evidence the answers cite, not that they determine the answer. The earlier figure of 81.3% for answers naming a Trustpilot or BBB rating counted any mention of either platform, and is dropped. The full change log is in the dataset.

## Data and downloads

- Every answer with its coded reputation and cited domains: [s16_answers_v11.csv](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_answers_v11.csv) and [JSON](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_answers_v11.json)
- All 5,188 claims with polarity, topic and attached sources: [s16_claims.csv](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_claims.csv) and [JSON](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_claims.json)
- The claim-support sample with both judges’ verdicts: [s16_claim_support.csv](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_claim_support.csv)
- Every cited domain and its source class: [s16_domain_classes.csv](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_domain_classes.csv)
- Every statistic on this page, including models and agreement: [stats.json](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/stats.json)
- Version 1.0 answers and statistics: [s16_legit_answers.csv](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/s16_legit_answers.csv) and [stats_v1.0.json](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/stats_v1.0.json)
- Machine-readable methodology: [methodology.json](https://underneath.agency/research-data/is-it-legit-ai-reputation-study/methodology.json)

The data is free to reuse with attribution (CC BY 4.0).

To cite: Underneath. (2026). *“Is this brand legit?” How AI assistants build a reputation* (Version 1.1). Underneath Research. https://underneath.agency/research/is-it-legit-ai-reputation-study

## Frequently asked questions

### Do AI assistants say a company is legit?

In our test, yes, every time. All 312 complete answers across ChatGPT, Gemini, Perplexity and Google AI Mode said the brand was legitimate. But 99.7% also made at least one negative claim, and 61.2% set out substantial concerns.

### What sources does ChatGPT use to check if a company is legit?

In our test, 92.4% of ChatGPT answers cited a review or complaint platform, most often the Better Business Bureau and Trustpilot. 70.9% cited the company’s own website. It rarely cited editorial review sites (7.6%) or news media (1.3%).

### Does Trustpilot affect what AI says about a business?

Trustpilot was the most-cited source, in 65.8% of answers, and claims citing only review platforms were negative 56.5% of the time. That shows Trustpilot is a prominent part of the evidence AI answers cite. It does not show that Trustpilot causes what the AI says; this study cannot test that.

### Are the complaints AI mentions accurate?

When the cited page could be read, it supported the claim fully or in part in 72.4% of sampled cases. The weaker point is frequency: 71.4% of claims that called a problem common cited evidence that did not show how common it was.

### Which AI assistant cites Reddit most for brand reputation?

Google AI Mode, in 40.5% of its answers, against 11.4% for ChatGPT. Adjusted for the brands asked about, its odds of citing Reddit were 5.76 times ChatGPT’s.

## Related research

- [Do ChatGPT, Gemini, Perplexity and Claude agree on brands?](https://underneath.agency/research/ai-assistants-brand-agreement-study)
- [Which Reddit threads do AI answers cite?](https://underneath.agency/research/ai-reddit-citations-study)
- [Ask an AI the same question 5 times: do the brands change?](https://underneath.agency/research/ai-recommendation-consistency-study)
- [Do AI answers match a business’s Google profile?](https://underneath.agency/research/ai-business-facts-accuracy-study)

## Related guides

- [Do Google AI Overviews downplay negative content?](https://underneath.agency/resources/do-ai-overviews-downplay-negative-content)
- [How often do AI answers say things their sources do not support?](https://underneath.agency/resources/ai-answers-unsupported-claims)
- [Should our reputation priorities for AI assistants differ from those for human customers?](https://underneath.agency/resources/ai-vs-human-reputation-priorities)
- [Can an AI search engine cite my page for something my page does not say?](https://underneath.agency/resources/ai-citing-pages-for-claims-they-do-not-make)
- [Do ChatGPT and other AI engines cite my own website or third-party reviews?](https://underneath.agency/resources/do-ai-engines-cite-your-own-website)

---

This is the Markdown twin of https://underneath.agency/research/is-it-legit-ai-reputation-study. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.