The claim that backfires
Almost every service website carries a sentence like it: the leading provider, the best agency, the number one in its region. In classic marketing, that sentence was mostly harmless. Nobody could hold it against you, and at worst it was ignored. In AI answers, that sentence has turned into a liability. You can watch it happen in prompt monitoring, the practice of asking AI assistants the same questions over and over and recording what they answer: when an assistant is asked for a recommendation, unbacked superlatives are not adopted. They are used differently.
Think of a colleague who only recommends a contractor after calling two references. An AI assistant works the same way, because every recommendation puts its own credibility on the line. Before it passes on a claim, it checks: does the website itself offer proof? Do independent sources say the same thing? A superlative that exists only as your own sentence fails that check. The assistant then has three ways to handle the failure:
The assistant simply builds its shortlist from providers whose claims it could verify. Your brand does not appear at all, and you never find out why.
Your claim is repeated, but wrapped in attribution: "describes itself as", "claims to be", "according to its own website". Every reader hears the asterisk.
Providers who back their claims typically get named with reasons: concrete case studies, named clients, verifiable credentials. The evidence becomes their pitch, delivered by the AI.
None of this required a competitor to attack you. The assistant took your own sentence, tested it, and turned the test result into the answer.
The distancing case is the one worth staring at. "X describes itself as the best agency in Germany" is technically accurate, sourced from your own website, and devastating in context, because it appears next to sentences like "Y is frequently recommended for Shopify projects and publishes detailed case studies." The assistant did not criticize you. It converted your unbacked claim into a credibility signal, with the sign flipped.
Test it yourself: two prompts and what to look for
You do not have to take any of this on faith. Open the AI assistant of your choice and ask a recommendation question in your own market. The question you type into the chat is called a prompt, and if you run an agency, these two work as templates:
Then read the answer like an auditor, not like a proud owner. Four things to check:
- Is your brand named at all? Absence is a finding, not a glitch. Note who appears instead.
- With what reasoning? Are you recommended for verifiable reasons, or merely mentioned? Reasons are what convince the user.
- With what wording? Watch for distance markers: "describes itself as", "claims", "according to their website". If they attach to you and not to competitors, that is your superlative problem in plain sight.
- Which sources does the AI cite? If the answer cites directories, review platforms and competitor case studies but not your site, the assistant found nothing on your site worth citing.
Run the same prompt again tomorrow and the answer will differ in detail. That is expected and it matters, but the pattern, who gets recommended with reasons and who gets distance wording, is far more stable than the individual answer.
The fix: back it or cut it
The repair rule is one sentence: every self-claim on your website either gets evidence or gets deleted. There is no third option, because an unbacked claim is not neutral. It actively invites the distancing treatment. Working through a website with that rule usually means three moves:
| Claim type | Unbacked version | Backed version |
|---|---|---|
| Quality claim | "The best agency for Shopify relaunches." | A case study per flagship project: initial situation, what was done, what changed, with the client named where possible. |
| Trust claim | "Hundreds of satisfied customers." | Testimonials with full names and companies, plus a link to a review profile the reader can open and count. |
| Authority claim | "Award-winning and certified." | The specific award and certification, named, dated and verifiable on the issuing organization's own site. |
Then make the evidence machine-readable, not just human-readable. That means the same discipline that makes any content citable: dedicated pages per case study rather than a slider of logos, testimonials as text rather than screenshots, claims phrased as checkable statements. How AI engines, the systems behind assistants like ChatGPT or Perplexity, select what they quote is covered in what AI actually cites; the short version is that an engine can only pass on what it can locate and verify.
Companies that do this consistently show up differently in AI answers, typically with the evidence itself as the recommendation reason. The case study becomes the sentence the assistant says about you. That is the full reversal: instead of the AI using your content against you, your content becomes the argument the AI makes for you.
From spot check to measurement
The two prompts above are a diagnosis, not a monitoring system. Single spot checks mislead in both directions: one lucky answer and you relax, one bad answer and you overreact. Engines differ, phrasings differ, and the same engine varies between runs. What you actually want to know is a trend: is my brand named more often this month than last? Did the distance wording disappear after we published the case studies? Which questions still go to competitors?
The stakes behind that trend line are not hypothetical. Visa expects millions of consumers to complete purchases through AI agents, programs that research and buy on a user's behalf, by the 2026 holiday season. And Gartner projects that by 2028, 90% of B2B buying, meaning companies buying from companies, will run through AI agents, at more than $15 trillion in purchase volume, as reported by Digital Commerce 360. Every one of those transactions starts with an answer like the one you just tested.
Answering that requires a fixed set of prompts, the questions your customers actually ask, run again and again across the different engines, with the answers recorded and compared over time. That is prompt monitoring, and it is what LLM Discovery Intelligence does systematically: same questions, every engine, every run, so a change in your AI visibility is a data point and not an anecdote. If your assistant-facing content also needs to stop inventing things in the other direction, that is the companion discipline described in how to prevent AI hallucinations.
One more thing your monitoring should cover: the sources you do not control. Assistants lean heavily on third-party voices when they weigh recommendations, and in a Semrush analysis across the major engines, Reddit was the most cited domain in AI answers (a domain is simply a website address). What relevant communities say about your brand, and whether it matches your claims, is its own trust signal, covered in depth in AI has to trust you before it recommends you.
Frequently asked questions
Why does an AI assistant not repeat my marketing claims?
Because an assistant is accountable to its user, not to you. Before it repeats a claim, it checks whether the claim can be verified anywhere: on your website, in directories, reviews or third-party sources. An unbacked superlative fails that check, so the assistant typically omits it, or repeats it with distancing wording such as "describes itself as", which signals to the user that the claim is yours alone.
What counts as evidence for a claim like "best agency"?
Anything a machine can verify: case studies with concrete, checkable details, customer testimonials with real names and companies, awards or certifications that exist on the awarding body's own website, consistent review profiles. The claim itself is not evidence. If nothing outside your own sentence supports it, an AI treats it as self-description, not as fact.
Is one test prompt enough to know where I stand?
No. A single answer is a snapshot: results vary between runs, engines and phrasings. One prompt is useful to see the mechanism with your own eyes. For decisions you need a fixed set of prompts, run repeatedly across several engines, so you can tell a real change from random variation. That is what prompt monitoring is for.
Should I remove all superlatives from my website?
Remove or back up, claim by claim. A superlative with evidence behind it, visible on your site and verifiable elsewhere, can be repeated by an assistant. A superlative without evidence is worse than silence, because it invites distancing wording and makes the rest of your content look less credible. The working rule: every self-claim either gets its evidence or gets cut.
Before you rewrite a single claim, find out whether AI engines can read and classify your website at all. The free check takes 60 seconds and needs no setup.