How should an AI agent decide which websites and recommendations to trust when researching a purchase?

Weight what costs money, action or time. Discount what costs nothing to write.

An agent researching a purchase should trust a source in proportion to what it would cost to fake. Text on a page, a review, a star rating and a stats dashboard cost nothing to produce, and with language models they cost nothing to produce at scale. Three things are still expensive: whether people actually read the site, whether the business shows money changing hands in records it does not control, and how long it has kept running. For recommendations from people, the expensive signal is a track record: how often their earlier tips worked out. Account age and karma are cheap by comparison, and on the data below they predicted almost nothing.

1. Check the seller against records it does not write

A shop's own site is a declaration. Public records are not. For a Norwegian business, the accounts in Regnskapsregisteret give revenue and Enhetsregisteret gives the founding date and headcount; Chrome's UX Report top lists say roughly how many people visit the domain; RDAP gives the domain's registration date. On Commit's index, read on 29 October 2026, komplett.no shows top 1,000 in Norway on reads, NOK 5.61 bn revenue in 2025, and a company founded in 1998 with 313 employees. A domain that does not exist reads as "not in the top 1M", no revenue and no registration. A shop whose row looks like the nonexistent domain on all three counts has told you nothing it could not have typed.

Our own row is on the same table: getcommit.dev is not in Chrome's global top 1M, which means too small to see, not unused. The signal has a floor. Below it, absence is not evidence of a fake.

2. Weight a recommender by past hits, not by followers

I tested this on 19 years of Hacker News: 113,483 YouTube videos first recommended there between 2010 and 2023. The 10% of picks from the highest-karma accounts caught on at 0.95 times the base rate. People whose earlier picks had caught on at least 10% of the time did 1.44 times better than people with no earlier hits (6.6% against 4.6%), and 1.54 times (1.33 to 1.78) with thread size held fixed. That is strangers recommending videos, not friends recommending products, and the effect sits in a small group: 3,657 videos from 789 people. It is still the cleanest public test I know of, and the code is public.

3. Treat reviews as declarations

Since August 2024 the US Federal Trade Commission's final rule on fake reviews bans reviews that misrepresent who wrote them, "such as AI-generated fake reviews", and bans buying reviews of a particular sentiment, with civil penalties for knowing violators. The FTC's guidance is the right place to read what is illegal. A rule raises the cost of faking; it does not let an agent tell a fake review from a real one by reading it. Prefer reviews tied to a verified purchase, and reviewers whose history you can see.

4. Know where the read signal goes blind

Traffic lists count humans. When we put the 121 work and payment platforms for agents in the Commonworld Atlas that have their own domain through the same row, 92 read exactly like the domain that does not exist on both Chrome's top-1M list and Cloudflare Radar. What did separate them was money: only 8 of the 121 appear as x402 sellers with settled on-chain volume. For agent-facing services, settled payments and counters read twice, weeks apart, say more than traffic.

A short checklist for the agent

  • Does the business show revenue or payments in a record it does not control?
  • How old is it, by registry or domain registration, not by its own "since" line?
  • Do people read the site, on a list the site cannot edit?
  • Is the recommender someone whose earlier tips you can check, and did they work out?
  • Which of the above did the source write about itself? Discount those.

Commit is ours, so weigh what this page says about it accordingly. The index figures were read on 29 October 2026 and the Hacker News figures come from the November 1–2 runs; both are on the linked pages with their sources.