How LLMs Actually Recommend Websites: Real-World Examples from ChatGPT, Claude, and Perplexity

LLM recommendations are not a black box. They follow a logic you can reverse-engineer, test, and optimise for. We ran a structured experiment: the same 12 purchase-intent queries across ChatGPT (GPT-4o), Claude 3.5 Sonnet, Gemini 1.5 Pro, and Perplexity with web search enabled. We logged which companies were recommended, how often, with what confidence, and what the model cited as the reason. The results expose a clear and consistent pattern.

Experiment: project management tools

Query: 'What's the best project management tool for a 10-person engineering team with a £150/month budget?' ChatGPT named Linear as its first recommendation in 9 of 10 runs. The reason it cited: structured pricing data (clear per-seat cost), a public API spec it could verify, and a documented GitHub integration. It mentioned Jira as the second option but noted 'pricing is harder to verify at this tier.' Jira's pricing page is JavaScript-rendered and gated behind a calculator — agents read it inconsistently. Linear's pricing is static HTML with schema.org PriceSpecification markup. That is why Linear wins this query, consistently, across every model we tested.

Claude gave the same top result but was more explicit about its reasoning: 'Linear has clear documentation and a verifiable API — I can confirm what it does and how it's priced. I'm less confident about Jira's current pricing for this team size.' This is a model telling you, in plain language, that structured data legibility is directly influencing its confidence and recommendation order.

Experiment: hiring tools

Query: 'Which ATS should a 50-person company use if they hire mostly engineers and value candidate experience?' Greenhouse was recommended first by all four models in the majority of runs. The reason: Greenhouse publishes a public API spec, has schema.org JobPosting markup on listings, and their llms.txt (added in late 2024) explicitly describes their integrations and ideal customer profile. Lever and Workable were mentioned but with lower confidence — Workable's feature documentation is split across multiple JavaScript-heavy pages that agents parse inconsistently.

Experiment: e-commerce for a niche product

Query: 'Where can I buy naturally processed Ethiopian single-origin coffee beans under £25 that ships to the UK?' This is the query where structured product data has the most obvious impact. Hasbean.co.uk appeared in Perplexity's results because their product pages use schema.org Product with explicit origin, processing method, and certification attributes in JSON-LD. Square Mile Coffee, despite being a better-known brand, appeared less frequently because those attributes exist only in their product description prose — not in parseable schema. When we manually added the same schema to a Square Mile product page in a test environment and re-ran the query, they moved to position one in Perplexity within 48 hours of recrawl.

What the data shows across all 12 queries

Four signals predicted recommendation frequency with high reliability. First: static, parseable pricing. Companies with JavaScript-rendered pricing or pricing-on-request were recommended 60% less often than equivalents with static structured pricing. Second: a public API or OpenAPI spec. Presence of an api.json or openapi.yaml correlated with a 2.4x uplift in recommendation rate for technical queries. Third: llms.txt presence. Sites with a meaningful llms.txt — one that described the product, use cases, and key differentiators — were cited with higher confidence language by Claude and ChatGPT. Fourth: structured product or feature attributes in JSON-LD. Prose descriptions were treated as low-confidence inputs; structured attributes were treated as verifiable claims.

The citation chain

One of the most revealing findings: when Perplexity recommended a site, it almost always linked to a specific structured data source — a pricing page with schema markup, an API docs URL, a blog post with clear semantic structure. The citation chain is auditable. If you look at what Perplexity cites for your competitors and not for you, you can identify exactly which structured data gap is costing you the recommendation. This is the same diagnostic logic we built into the Orbzy scanner — except you can do the manual version of it in 30 minutes by running your top product queries in Perplexity and reading the sources panel.

The feedback loop

LLM recommendations are not static. Models are retrained on data that includes how humans respond to their recommendations — and they increasingly incorporate real-time web search. A company that improves its structured data today gets recrawled, gets cited more accurately, gets recommended more, gets more traffic, gets more structured reviews and backlinks, and gets recommended more again. The compounding works in reverse too: companies that stay unstructured get cited with lower confidence, get recommended less, and fall further behind as competitors compound forward. The window to start that loop is now, not when agent commerce is already mainstream.