How much an AI agent saves checking a business
We asked 12 AI agents the same question about the same businesses: is it legit, and what are the facts? Four read the Truth Signal report. The other eight went looking on their own. Reading the report took about 30 seconds, 7 times fewer tokens and 26 times fewer tool calls, and answered every fact. Going it alone took about 3 minutes and still left 32% of the facts blank.
The job we gave them
Each agent was asked about one of two real businesses: johnlewis.com, a large UK retailer with a long history, and designbundles.net, a smaller online business. Each was given the same list of facts a careful person would want before spending money: the company on the register, its number and status, its data-protection registration and VAT number, how old its website address is, when the web archive first saw it, its contact details and returns policy, its reviews on every platform, its followers on six networks, any data breaches or malware warnings, its email protection and its certificate. Then a trust score and a verdict. One rule: if you cannot find a fact, say so; never invent one.
Three ways to do it
Four agents could read only the business’s Truth Signal report, as Markdown. Six could search the web and read any page they liked. Two could do all of that and run their own lookups from a terminal, as a developer’s agent can. We used three Claude models, Haiku, Sonnet and Opus, from the smallest and fastest to the strongest.
What reading the report saved
The agents that read the report made 1 or 2 tool calls and finished in about 30 seconds. The agents that went looking made 32 on average, up to 54, and took about 3 minutes, up to 5. Every turn an agent takes, it reads its whole conversation again, so the tokens it processes grow with every step: about 225,000 for an agent reading the report, about 1.6 million for one going it alone. That is 7 times fewer tokens, 6 times faster and 26 times fewer tool calls.
And they found less
The report agents answered every fact. The agents that went looking left 32% of the facts blank and got three wrong, and the best of them found 88%. Follower counts were the hardest: the networks keep their pages from automated visitors, so every agent that went looking left at least half of them blank. Google review figures, VAT numbers and the technical facts (email protection, the certificate, the age of the website address, the first archive capture) came next, because they need a lookup no ordinary web page shows.
Why the gap is so large
A report has already read the company register, the data-protection register, the domain records, the web archive, the review platforms, the social profiles, the breach and malware lists and the email settings, and kept each fact with its source. An agent on its own has to find each of those places, learn how each one answers, and read pages built for people, with their menus, adverts and cookie banners. Each page it reads stays in its memory for the rest of the job, so the cost of every later step grows.
A stronger model is not the answer
The strongest model, Opus, took about 4 minutes and 2.3 million tokens a business when it went looking, and still left facts blank. The smallest, Haiku, reading the report answered every fact in about 36 seconds. The cheaper models that went looking gave up early: fast, but with many facts blank.
How we measured it
Each agent ran on its own, with no memory of the others. We counted every tool call, the time from its first step to its last, and every token it processed: everything it read on each turn, including what it read again from its memory, and everything it wrote. We compared each answer with the official sources fact by fact. Where an answer was right in more than one way, such as a business with two phone numbers, or a social media count that had grown since our report, we set that fact aside for every agent. This is two businesses and one run per agent: it shows the size of the gap, not a precise rate. Every agent’s figures:
| Business | Model | Given | Facts right | Wrong | Tool calls | Time | Tokens |
|---|---|---|---|---|---|---|---|
| johnlewis.com | Haiku | The report | 26 of 26 | 0 | 1 | 0:30 | 208,111 |
| johnlewis.com | Sonnet | The report | 26 of 26 | 0 | 1 | 0:25 | 205,975 |
| designbundles.net | Haiku | The report | 25 of 25 | 0 | 2 | 0:43 | 281,567 |
| designbundles.net | Sonnet | The report | 25 of 25 | 0 | 1 | 0:24 | 205,623 |
| johnlewis.com | Haiku | The web | 18 of 26 | 1 | 37 | 3:17 | 1,688,115 |
| johnlewis.com | Sonnet | The web | 11 of 26 | 0 | 8 | 0:35 | 357,425 |
| johnlewis.com | Opus | The web | 21 of 26 | 1 | 54 | 4:43 | 2,460,954 |
| johnlewis.com | Opus | Web and terminal | 22 of 26 | 0 | 36 | 4:19 | 2,676,106 |
| designbundles.net | Haiku | The web | 14 of 25 | 1 | 31 | 2:30 | 1,049,565 |
| designbundles.net | Sonnet | The web | 9 of 25 | 0 | 8 | 0:41 | 425,894 |
| designbundles.net | Opus | The web | 19 of 25 | 0 | 45 | 3:33 | 1,604,270 |
| designbundles.net | Opus | Web and terminal | 22 of 25 | 0 | 39 | 3:53 | 2,373,460 |
The quick checklist
- Fetch a report as JSON at
/business/{domain}.json, or as Markdown at/business/{domain}/report.md. - In ChatGPT or Claude, add
https://truthsignal.com/mcpas a connector: no sign-in, no key. - Point a crawler or a coding agent at llms.txt.
- Read the facts, not just the score: each one says where it came from.
- Quote the truth score with a link to the report, so the reader can see every source.
- A business nobody has asked about is checked when you ask; ask again when the answer says to.
Questions people ask
How does an AI agent use Truth Signal?
It reads one report per business: as a web page, as JSON or as Markdown, or through our MCP connector in ChatGPT or Claude. It is free, with no key and no sign-up. Our setup guide and llms.txt say how.
Why not let the agent search the web itself?
It can, but in our test that took about 3 minutes and 32 tool calls a business, and still left 32% of the facts blank. Many of the facts sit in registers and records an ordinary web page does not show.
Does a stronger model make up the difference?
Not in our test. The strongest model going alone used about 2.3 million tokens a business and still missed facts; the smallest model reading the report answered every one.
Does it save money?
Fewer tokens usually means a smaller bill, but much of what a searching agent reads is read again from its memory, which is charged at a lower price. So we measured tokens, time and tool calls, and make no claim about cost.