---
title: "What AI Actually Reads: ChatGPT, Gemini, Perplexity and Google AI Mode Compared"
description: "What ChatGPT, Gemini, Perplexity and Google AI Mode actually read before recommending a firm - the study-backed, platform-by-platform breakdown."
url: https://www.agiledigitalagency.com/blog/what-ai-reads-on-your-website/
date: 2026-08-25
modified: 2026-08-25
author: "Agile Agency"
image: https://www.agiledigitalagency.com/wp-content/uploads/2026/08/what-ai-reads-on-your-website.avif
type: blog
lang: en
---

# What AI Actually Reads: ChatGPT, Gemini, Perplexity and Google AI Mode Compared

By the time an AI assistant recommends a firm – a solicitor, an accountant, a financial adviser, a consultancy – it has already done its reading. The recommendation is the last step of a process that started with sources: pages fetched, profiles checked, reviews weighed. Most firms have no idea what that reading list contains, which makes “how do we get recommended?” feel unanswerable.

It is more answerable than it looks. Over the past two years, several large studies have tracked exactly which sources the major AI engines cite when they answer buying-intent questions about businesses – and the findings are specific enough to act on. The clearest lesson is that each engine reads differently, and sometimes in flatly opposite ways: the review platform one engine consults in every industry is the one another never cites at all. This is the evidence-based comparison – platform by platform, with the numbers and their limits.

**A scope note before the numbers.** Most platform-level studies cited here examine local-business searches rather than professional-services firms specifically. The practical lessons carry over to law, finance and consulting – but the exact source mix may differ by sector, geography and query type.

One term needs pinning down too: in this article, a “source” means a page, listing or domain observed in a study’s output – not a disclosed retrieval log. Being crawled, being retrieved into an answer, appearing as a source, being mentioned by name and being recommended are five different events, and the studies here mostly observe the middle three.

## In this article:

- [Platform by platform: what each engine actually reads](#platform-by-platform)
- [The common thread – and why it isn’t a rule](#four-engines-four-reading-lists)
- [The three surfaces a professional-services firm controls](#the-three-surfaces)
- [The on-site checklist: what AI reads on your website](#what-ai-reads-on-your-website)
- [What this means sector by sector](#sector-by-sector)
- [What these studies can and can’t tell you](#what-these-studies-cant-tell-you)
- [Build your own baseline](#build-your-own-baseline)
- [FAQ](#faq)

## Platform by platform: what each engine actually reads

This table is the heart of the piece. It synthesises what the tracked studies observed each engine actually consulting when answering buying-intent questions about businesses – each row limited to what the named study recorded:

| Platform | Most prominent observed source pattern | Standout observed finding |
| --- | --- | --- |
| **ChatGPT / ChatGPT Search** | Business websites frequently observed; then mentions of the business on other sites; then directories | Business websites used 58% of the time, business mentions 27%, online directories 15% (BrightLocal, Dec 2024). In the July 2025 study, dental queries were sourced exclusively from ten dental directories. |
| **Gemini** | Business websites; selective with third-party review platforms | Across every industry tested, Gemini never cited Yelp directly (BrightLocal, July 2025). |
| **Perplexity** | Business websites plus review platforms, cited openly | Used Yelp as a source in every industry tested – the heaviest review-platform reliance of the four (BrightLocal, July 2025). |
| **Google AI Mode** | Google’s wider search and local ecosystem alongside the website | Google Business Profile appears to be an important information source for Google’s local AI surfaces (BrightLocal, July 2025). |

A few of those rows deserve unpacking. The ChatGPT figures come from [BrightLocal’s December 2024 analysis of ChatGPT’s search sources](https://www.brightlocal.com/research/uncovering-chatgpt-search-sources/): business websites 58%, business mentions 27%, directories 15%. That is a strikingly website-first diet – but the July 2025 study showed how sharply it can vary by sector, with dental queries answered exclusively from ten dental directories. Whether your profession has an equivalent set of gatekeeping directories is worth finding out before your competitors do.

The Yelp split is the clearest illustration that the engines genuinely read differently. Across the whole July 2025 study, Yelp appeared as a source in 33% of all searches – yet that average conceals two opposite behaviours: Perplexity used Yelp in every industry tested, while Gemini never cited it directly at all. The same review page can be a primary source on one platform and invisible on another. And because Google’s AI surfaces draw on the wider search and local ecosystem, the Google Business Profile appears to be an important information source for Google’s local AI answers – which makes your GBP closer to source material than an optional listing.

## The common thread – and why it isn’t a rule

The table’s differences sit on top of one consistent pattern. In July 2025, [BrightLocal ran the same local-business searches across four AI platforms](https://www.brightlocal.com/blog/ai-search-using-listings-sources/) – Google AI Mode, Gemini, Perplexity and ChatGPT Search – twenty searches in each of ten business niches, and recorded every source each platform cited. Their summary finding:

**“The vast majority of sources across every single LLM and industry were businesses’ own websites.”** – BrightLocal, July 2025, across 20 searches × 10 niches × 4 AI platforms.

Held alongside the rest of the evidence, the honest version of that conclusion is this: across the studies we’ve reviewed, a business’s own website is consistently one of the most important source types – but the balance changes dramatically by platform, industry and query. The dental example above is the proof of the variance: in that sector, on that platform, in that month, the businesses’ own websites did not lead at all – ten directories did. The practical consequence still stands, just without the absolutism: a site that is thin, ambiguous or unreadable to machines weakens you across every platform simultaneously, because it is the one source type they all draw on heavily – even if none of them draws on it exclusively.

## The three surfaces a professional-services firm controls

Two further studies complete the picture for professional services specifically. [Profound, whose index covers more than 30 million AI citations](https://www.tryprofound.com/blog/linkedin-is-the-most-cited-domain-for-professional-queries-in-ai-search), found that for professional queries, LinkedIn is the number-one most-cited domain across all six major AI platforms. And the Trustpilot and Seer Interactive analysis of over 800,000 AI answers found actively-reviewed brands cited in 75.3% of relevant answers against roughly 1% for brands without an active review presence – an observational association, not a controlled experiment, which we unpack in [the pillar on website mistakes](/blog/website-mistakes-ai-search/).

Put the studies together and a professional-services firm’s AI presence rests on three surfaces: **your website**, the source type every engine draws on heavily; **your LinkedIn presence**, the most-cited domain for exactly the kind of queries your clients ask; and **your review footprint**, which the Trustpilot data suggests functions almost as an entry requirement. Only the first is fully yours. LinkedIn and the review platforms set their own rules, rank by their own logic, and can change beneath you – your website is the one surface where you decide every word the engines read. Those three deserve priority, but they are not the whole footprint: professional-services firms should also consider regulatory registers, professional memberships, directories such as Chambers and Legal 500, industry associations and press coverage – third-party sources that corroborate identity and expertise. That is also why it deserves the most scrutiny: when your own site says too little, engines assemble their answer from third-party sources instead, with all the risks of [ghost citations](/blog/ghost-citations-ai-search/) that follow.

## The on-site checklist: what AI reads on your website

If the website is the common thread, what exactly should be on it? [BrightLocal’s guidance on what AI looks for on a business website](https://www.brightlocal.com/learn/ai-and-local-search-tips/) is the most concrete published checklist, and it maps cleanly onto professional services:

- **Name, address and phone on the homepage and contact page** – stated in text an engine can parse, not buried in an image or a footer script. For a firm, consistency with your LinkedIn and directory listings matters as much as presence.
- **One page per service.** A single “our services” list gives an engine one thin source; a page per practice area gives it a citable answer for each question a client might ask.
- **Opening hours** – mundane, and exactly the kind of factual detail an assistant checks before recommending somewhere a person can actually reach.
- **FAQ content** that answers real client questions in full sentences. AI answers are assembled from passages that already read like answers.
- **Reviews linked or surfaced on the site**, connecting your strongest third-party signal to a source every engine reads.
- **An about page that tells the firm’s actual story** – who you are, who you act for, what you are known for. Entity clarity starts here.
- **Real photos** of people and premises rather than stock imagery.
- **First-hand experience content** – matters you have handled, judgements you have formed – the material no competitor and no language model can generate about you.
- **HTTPS and mobile speed**, because a page that cannot be fetched cleanly cannot be read at all.

In our own [generative engine optimisation](/services/generative-engine-optimization-geo/) work we treat that list as the floor, not the ceiling – layering on structured data, citable content structure and crawler-level readability. But the floor matters: most professional-services sites we audit fail several of these basics before any advanced work begins, and the most common failures are catalogued in our guide to [the website mistakes that keep firms out of AI search](/blog/website-mistakes-ai-search/).

## What this means sector by sector

The checklist is generic by design; what each firm needs to expose is not. Applying the findings to the sectors we work with:

| Sector | What the website needs to expose |
| --- | --- |
| **Law** | Practice areas as individual pages, the jurisdictions you act in, solicitor profiles with credentials, and representative matters (anonymised where needed). |
| **Financial services** | Authorisation status and regulator, specialisms, the required risk warnings, and team credentials an engine can verify against the register. |
| **Consulting** | Sector experience, named consultants, the methods you actually use, and measurable outcomes rather than adjectives. |
| **SaaS** | Capabilities stated plainly, integrations, customer evidence, and public documentation an engine can quote. |

## What these studies can and can’t tell you

A word of honesty before the conclusions harden into rules, because the samples behind these figures are worth stating plainly. BrightLocal’s July 2025 study ran 20 searches in each of 10 niches per platform – rigorous for its purpose, but 200 searches per platform is a sample, not a census, and a niche it did not test may behave like dentistry rather than like the average. The December 2024 ChatGPT figures come from a separate, earlier sample of the platform’s search behaviour. The Trustpilot × Seer finding is observational, as noted above. And every figure here shares three structural limits: it is tool-observed (what a tracking setup recorded, not the platform’s own disclosure), it is point-in-time (particular queries, particular months), and the platforms change their weightings without notice – the engine that ignores a directory this quarter may lean on it next. What the studies establish reliably is the shape of the behaviour: engines draw heavily on business websites, they diverge sharply on third-party sources, and reviewed, well-documented firms appear vastly more often than invisible ones. Treat these findings as a current working picture, not a permanent technical specification. Build for the shape, then [measure your own appearances](/blog/ai-visibility-score/) rather than trusting anyone’s averages – including these.

## Build your own baseline

The measuring is simpler than it sounds, and worth doing before any optimisation:

1. Select 15-25 buyer-intent questions your prospective clients actually ask.
2. Test them across the platforms relevant to your market.
3. Record four outcomes separately: retrieved, cited, mentioned, recommended.
4. Split the results branded vs non-branded, and informational vs comparison.
5. Repeat monthly, so change shows up as a trend rather than an anecdote.

The [AI Visibility Score](/blog/ai-visibility-score/) is the systematic version of exactly this exercise – a fixed prompt set, scored engine by engine, period over period.

## FAQ

### What sources does ChatGPT use when recommending businesses?

BrightLocal’s December 2024 analysis found ChatGPT drew on business websites 58% of the time, mentions of the business on other sites 27%, and online directories 15%. Sector matters, though: in BrightLocal’s July 2025 study, dental queries were sourced exclusively from ten dental directories. These are tool-observed findings from specific samples, not fixed rules.

### Do AI engines read review sites?

Yes – but very unevenly. In BrightLocal’s July 2025 study, Yelp appeared as a source in 33% of all searches: Perplexity used it in every industry tested, while Gemini never cited it directly. Review activity is also strongly associated with appearing in AI answers at all – the evidence for that is covered in [our pillar on the website mistakes that keep firms out of AI search](/blog/website-mistakes-ai-search/).

### What should a professional-services firm put on its website for AI search?

The essentials BrightLocal’s guidance identifies: name, address and phone on the homepage and contact page; a dedicated page per service; opening hours; FAQ content; reviews linked from the site; a genuine about-page story; real photos; first-hand experience content; and HTTPS with good mobile speed. Beyond that floor, structured data and citable content structure – the substance of generative engine optimisation – make each page easier for engines to quote.

### What do the engines read – and say – about your firm?

If you want to know what the engines currently read – and say – about your firm specifically, our [SEO & Marketing Intelligence Report](/services/seo/seo-audit/) audits all three surfaces against the platform behaviours above.