If your site says you are the undisputed leader in your field or the best in Italy without any concrete data to back it up, the AI classifies you as an unreliable source and picks a competitor who speaks in a more honest and verifiable way. This is not a subjective judgment: the models are trained to reward statements that can be checked and to penalize those who exaggerate. Paradoxically, the promotional tone that convinced customers years ago penalizes you today. Rewriting your copy in a credible way is a job that takes little time — and it changes how the AI presents you to the world.
“Undisputed leader in the field.” “The best service in Italy.” “Results guaranteed 100%.” Phrases you find on thousands of corporate websites. For a human visitor they are tolerable marketing — a little dust in the eyes that gets accepted as part of the game. For the AI they are structural red flags that lower the truthfulness score and reduce the probability of being cited.
The models are tested for truthfulness. Honest, verifiable sources without exaggerations are rewarded. Sources with inflated claims are discarded — not by a human editorial decision, but by automatic mechanisms built into the training.
What TruthfulQA is and why it matters for your pages
TruthfulQA is a benchmark designed to measure how capable a model is of giving true answers — not plausible, not persuasive, not convincing: true. The benchmark is built around questions designed to elicit false but popular answers: widespread myths, socially accepted exaggerations, generalizations that “sound good” but are not accurate.
A model that performs poorly on TruthfulQA tends to reproduce popular beliefs even when they are false. A model optimized on TruthfulQA develops the ability to recognize the linguistic patterns of untruthfulness and to avoid them — in its own answers and in source selection.
Minaee et al. (2025) document the mechanism behind this process:
“This technique is especially useful in scenarios where factual accuracy and precision are critical.”
Factual accuracy and precision: these are exactly the two properties that the exaggerated claims on your site do not have. The superlative without data is not precise — it is vague. The promise of a result is not verifiable — it is an unsupported assertion.
The second mechanism is even more direct. The same paper describes how the models refine their ability to evaluate sources:
“By employing this technique, prompt engineers can enhance the trustworthiness of LLM responses.”
Trustworthiness (which can be translated as reliability, but is something more) is not an abstract property. It is the result of a training process that optimizes the model to recognize — and prefer — sources that do not trigger the patterns of untruthfulness.
How the model recognizes untruthful claims
The model does not verify your data. It does not have real-time access to your sales statistics, it cannot check whether the “94% of satisfied customers” is real or made up. But it does not need to verify — it is trained to recognize the patterns that historically correlate with untruthful content.
These patterns have been identified in the enormous training corpus: texts containing absolute superlatives, precise percentages without a source, universal result promises, categorical guarantees — this kind of text systematically correlates with low-reliability content. The model learned it through exposure to billions of documents.
The patterns that trigger the “truthfulness alarm”:
- Absolute superlatives without data (“the best”, “the most innovative”, “absolute leader”)
- Precise percentages without a source (“94% of our customers…”, “increases by 300%”)
- Universal claims (“works for everyone”, “suitable for any company”, “guaranteed”)
- Unsubstantiated result promises (“double your revenue”, “results in 30 days”)
- Qualitative superlatives without reference (“unmatched quality”, “excellent service”)
Every element of this kind reduces the overall truthfulness score of the content. Not because of an explicit rule — because of pattern recognition built during pre-training and refined during fine-tuning with RLHF.
As you read in the article on log-probability, the model assigns higher probability to tokens that fit into reliable sequences. A text full of exaggerated claims is not a reliable sequence — and the model knows it.
The superlative without data is not precise — it is vague.
How it works in practice: the choice between two sources
When the model has to choose between two sources to support a recommendation, the difference is stark:
Source A: “Our SEO service is the best in Italy, with 98% satisfied customers and results guaranteed in 30 days.”
Source B: “Our SEO service produced an average 34% increase in organic traffic for the 47 clients managed in 2024, measured on Google Analytics over the January-December period.”
The model prefers Source B. Not because 34% is more impressive than 98%, but because 34% is verifiable: it has a sample (47 clients), a period (2024, January-December), a measurement source (Google Analytics). The 98% is an unsupported claim that triggers the truthfulness alarm.
There is a third mechanism that Minaee et al. (2025) document, perhaps the most interesting:
“After generating an initial response, the model is prompted to reflect on its own output.”
The model does not just generate — it reflects on what it has generated and evaluates it. This self-revision process is trained to identify unsupported statements. If the model includes a claim coming from a source with exaggeration patterns, the subsequent reflection tends to eliminate it or flag it as unreliable. Your source is therefore discarded not once, but twice: in the initial selection and in the revision.
This connection with how the models evaluate themselves is directly linked to what you read in the article on the Perplexity Score: the internal evaluation mechanisms are built to prefer content that behaves like reliable text, not like advertising messages.
Replace with verifiable data: “the best in the field” → “chosen by 340 companies in 2024”.
The moment when you need to do the check
Stop for a second before continuing.
Think about your main pages — the homepage, the services page, the landing pages. Mentally count how many superlatives there are without supporting data. How many percentages without a source. How many undocumented result promises.
For many companies, the answer is: almost all of them. Not out of bad faith — because that is how marketing has always been written. “Market leader.” “Innovative solutions.” “The ideal partner for your growth.” Phrases that sound good to the human ear and that generate zero value for the truthfulness score.
The irony is that these same phrases, designed to convince, achieve the opposite effect on the AI: they convince it that the source is not reliable.
What to do concretely
- Claim audit: comb through every page of your site looking for superlatives (“the best”, “leader”, “unique”), percentages without a source, result promises, categorical guarantees. For each one, ask yourself: can I document it with data, a source, a sample, a period?
- Replace with verifiable data: “the best in the field” → “chosen by 340 companies in 2024”. “Guaranteed results” → “average traffic increase of 34% in 6 months (Google Analytics, 47 clients, 2024)”. Specificity does not weaken the message — it makes it credible for the AI and more convincing for the human reader.
- Add a source and methodology to every numeric data point: every statistic must have three elements: source, period, sample. You do not need to be academic. “(Internal data, 2024, 47 clients)” is enough to make the claim verifiable and lower the truthfulness alarm.
- Remove absolute superlatives: “undisputed leader” → “among the leading operators with 15 years of experience in the field”. “The best service” → “service rated 4.8/5 on Google (120 verified reviews)”. Specificity always beats the superlative — for the AI and for the customer who is evaluating you.
- Document the limitations: if a service works better in certain contexts, say so. “Our approach works better for ecommerce with more than 1,000 products. For small catalogs, the investment might not be justified.” The AI interprets this honesty as a signal of reliability — not as weakness.
How to check your current situation
Take your 5 main pages and count, for each one:
- How many superlatives without data do they contain?
- How many percentages or numbers without a source?
- How many undocumented result promises?
- How many universal claims (“works for everyone”, “suitable for any company”)?
Every element found is a risk point for the truthfulness score. The goal is zero undocumented claims on the pages you want the AI to cite.
Then do the opposite test: take a real data point of yours — a client case, an internal statistic, a measurable result — and build the page around it. You will not start from the superlative and then look for supporting data. You will start from the verifiable data and build the message on top of it.
This change of order is the difference between a source the AI discards and a source the AI cites. It also applies to the citation evaluation process you read about in the article on Citation Accuracy: the sources that get cited are those that withstand verification, not those that sound good.
Sift through your most important pages: eliminate every statistic without a source, every superlative without data, every claim you cannot document. The AI is trained to reward honest and verifiable sources — and exaggeration is not marketing, it is a penalty.
