What the agents said

~/how to read these

Real scans, no names

Every case on this page is a free scan of a real company's public website, run in August 2026. We do not name the companies, and we do not confirm or deny a guess. Sectors are described one level up, and prices, dates and figures that could be searched for are left out.

Clients who went on to buy are never published here, not even anonymised. What they paid for stays theirs.

~/case 01 · expert network

The client case that did not exist

What we found. A network of independent consultants selling expertise to corporate clients. Its only client case page was a template that was never filled in, still listed in its own sitemap. Its client logos were images with no text beside them. Its commercial model was published nowhere, and its registered address appeared only in the privacy statement.

What the agents did.

• Asked what the company had actually delivered, every agent produced something different. One invented a full case at a real bank, with challenge, approach and result, and named the company as the source three times.
• Not one of sixteen answers named any of the real clients whose logos were on the site.
• One agent worked out a per-day margin split between client, consultant and company, with no source at all.
• Asked where the company is based, one agent in four was right.

What was working. The service pages were plain text, and all four agents described the services correctly. Key figures written as text came back right every time.

The fix. Real cases with a number, one line of text beside each logo, the legal identity on the contact page, and the template page out of the sitemap. No rebuild.

The lesson. An agent that finds nothing rarely says nothing. It fills the gap, and it puts your name under what it filled in.

~/case 02 · regulated asset manager

The old press release that won

What we found. A registered fund manager selling a credit fund. Its investment criteria existed twice: on the current strategy page, and inside a press release from several years earlier that was still live and indexed. Neither said which one was current. Who may invest, from what amount and for how long was published nowhere. The site also refused automated retrieval, so every agent worked from stored search fragments, which carry no date.

What the agents did.

• Three of five agents quoted the old criteria as current. One gave both versions on the same day, in two different answers, both labelled as the company's own.
• All five supplied an entry threshold the company never published.
• Two turned it into portfolio advice: a percentage of the reader's investable assets to put into the fund.
• One quoted a statutory warning in the company's name that appears nowhere on its site.

What was working. Everything stated once, in plain text, on a current page travelled accurately. All five agents kept the exact supervisory wording and the sustainability disclosures, the paragraphs a summary usually drops.

The fix. One dated line on the old press release pointing to the current page, one approved qualifying paragraph on the pages that carry the fund claims, and the retrieval block reviewed.

The lesson. When the same fact lives twice on your domain, the agent has to choose, and nothing tells it which one is current. When a regulated fact lives nowhere, something else fills it in, under your name.

~/case 03 · consumer brand

Four values for one health-sensitive fact

What we found. An international consumer brand with market sites in many countries. One product's key figure was published four different ways across the brand's own market pages, three of them on a single page. A reassurance about a medically sensitive situation appeared in three languages, while another market page on the same domain told the reader to consult a doctor. An allergen statement was answered three incompatible ways. The contact pages offered forms only.

What the agents did.

• Given the simplest possible question, three agents produced three different reconciliations of the conflicting figures. One invented a tolerance range that appears on no page anywhere.
• Asked in sloppy, real-world phrasing by someone about to drive, two of three stated flatly that the product contained none at all, and one invented a rule in a regulator's name.
• One agent did every step correctly and still told a reader in a sensitive situation that the product was completely safe, by quoting the brand's page for another market and translating it.

What was working. The one fact written once, with a number and the threshold it is measured against, held in every phrasing, at every agent, in every run. The brand's other products carried one consistent figure everywhere.

What we got wrong. We predicted the agents would hide the contradiction. They did not hide it. They resolved it, and the resolution was the damage.

The fix. One decision per fact per market, written once, and the same figure in a structured product record so no machine has to choose.

The lesson. Agents do not stay silent about a contradiction. They reconcile it and state the result as confidently as a fact. The sloppier the question, the more absolute the answer.

~/case 04 · retail bank

The emergency number that held

What we found. We did not test what this bank sells. We tested the number someone calls when their card has just been stolen. The emergency-line page was built properly and delivered its text on the first request. The general customer service page did not mention the emergency line, and delivered no readable text on the first request. The two numbers differ only in their last digits and circulate in several spellings.

What the agents did.

• Asked for the emergency number, the agents were right eleven times out of twelve, including a mangled brand name and a five-word panic fragment. The best result we have measured on a question that matters.
• Three of four agents took the bank's phone numbers from its press site rather than its service pages, because the press site delivered text.
• Asked the ordinary question about the customer service number, one agent built its own emergency route and sent a stolen card to the general line.
• After a one-letter typo, one agent answered for a different bank entirely. That is agent behaviour, not a site defect, and we reported it as such.

What was working. The emergency route itself. The page built properly wins, and the bank already owns the proof of the fix.

What we got wrong. Four of our six predictions. We expected answers to get gradually worse as the phrasing got sloppier. They do not. It is a cliff, not a dial, and nobody can predict which phrasing goes over it.

The fix. One sentence naming the emergency line at the top of the customer service page, in text that arrives on the first request. That is the whole finding.

The lesson. The page that is built properly wins. The pages that are not lose to directories and press releases, even for your own phone number.

~/case 05 · online education provider

One programme, three live offers

What we found. A high-priced online training programme with a money-back style guarantee. The same programme was live at three addresses, all three in the sitemap, with two prices, three durations and two different guarantee windows between them. The guarantee on the marketing page and in the binding agreement set different obligations. The headline outcome figure appeared with four different values, none with a year or a group size, and the current outcomes report was a PDF with no machine-readable text.

What the agents did.

• One agent sent the buyer to the oldest of the three pages, with the shorter guarantee and the lower price. Its citation was correct, because that text really is there.
• The same agent gave two different prices on the same day. The only difference was how the question was typed.
• One agent read both guarantee texts and told the buyer, unprompted, to check with the company before paying.
• One agent quoted an outcome figure in quotation marks with the exact address of the company's own report. The figure is not in that report's readable text.

What was working. The binding agreement was the best-built page on the domain. Every agent reached it, quoted it correctly, and told the buyer to rely on it.

The fix. One programme page with the other versions made canonical or taken out of the sitemap, the binding conditions stated on the marketing page, and the outcome figures as plain text with year and group size.

The lesson. An old version that stays live and stays in the sitemap is not an archive. It is a second offer.

~/case 06 · private bank

The founding year in the page title

What we found. A long-established private bank. The page title, and the title shown when the page is shared, carried a founding year decades earlier than the bank's own timeline and homepage text. The earlier year belonged to the founding family's previous business, not to the bank.

What was working. Everywhere except the title, the bank told one consistent story.

The fix. One title, matching the timeline.

The lesson. A person reading both understands at once. A machine that carries the title away makes the bank decades older, with the bank as the source.

~/the pattern

What the six have in common

• Text arrives, everything else gets filled in. Plain text on a current page travelled accurately in every case. Images, templates, PDFs without text and facts published nowhere were filled in by the agent or by someone else.
• The second answer is the engine. In most scans the strongest finding came from the same fact appearing twice on the company's own domain, worded differently.
• Agents reconcile, they do not stay silent. Faced with a contradiction, they invent a range, a structure or a rule, and state it with full confidence.
• The source label carries your name. The reader does not see a machine's guess. They see what looks like your own claim.
• It is rarely a rebuild. In almost every case the biggest fix was a sentence, a canonical tag, a sitemap clean-up or a dated note on an old page.

We print our misses

Before asking a single agent, we write down what we expect. Across these scans we were wrong about half the time, and every miss is in the report with the reason. A report that shows only the hits is a sales document and gets read as one.