Figures, dates and sources: making content verifiable for an AI.

A model does not verify a claim, it cross-checks it

Content that is verifiable for an AI is content in which every important claim carries what it needs to be cross-checked: a precise figure with its unit and its scope, a validity date, a named source or a described method. The model does not “verify” in the human sense. It does two things: it detects the presence of these markers, and it compares the claim with what other sources say on the same question. A marked, concordant claim is retained; a bare or isolated claim is an opinion.

This behaviour is documented only in broad outline: the vendors describe a selection of sources based on relevance and perceived reliability, without publishing their criteria. What we describe here therefore belongs to two registers, what is publicly documented and what we observe on our client engagements by comparing cited passages with ignored ones. We distinguish the two every time.

Verifiability is the part of the second condition of a citation, “citable”, that concerns proof. It also prepares the fourth, “corroborated”: a precise, dated claim is the only kind a third-party source can repeat word for word. The method article sets out these four conditions; this one details the proof.

The four markers of a verifiable claim

In the passages we see picked up by the engines, four markers keep recurring. None is sufficient on its own; it is their accumulation that separates a cited sentence from an ignored one.

The fourth marker is the least practised and, in our observations, the one that most often decides between two sources with differing figures. Between “the average time is seven weeks” and “across the 38 sales we handled in 2025, the median time between listing and preliminary sale agreement was seven weeks”, the second wording is more modest and more often picked up, because it says what it measures.

Figures: precision, unit, scope

A citable figure has a value, a unit, a scope and a base. “30%” has only the value. “30% savings on the heating bill” has the unit and the object. “30% savings on the heating bill of a 120 m² gas-heated house, compared with the previous season, across the twelve households monitored in 2025” has everything, and it is this version the model can use without wrongly generalising it.

Precision is not false precision. A figure to one decimal place on twelve cases is less credible than a range. A range with its bounds and its condition (“9 to 21% depending on the condition of the old frames”) is picked up as readily as a single figure, often as is, because it contains its own nuance.

The percentage without a base is the most frequent flaw. “95% satisfied customers” says neither how many answered, nor which question, nor when. “95% of the 214 respondents to our March 2026 survey say they are satisfied with the installation” says everything, in ten more words.

Dates: date the claim, not just the page

The date that matters to a model is the one at which the claim is true, and it is rarely the page date. A page updated in September 2026 that quotes a “current” rate without dating it leaves the model guessing. A claim dated in the text (“as at 1 September 2026”, “in the first half of 2026”) can be cross-checked and expires cleanly.

Three dates coexist on a page. The publication and modification dates, carried by datePublished and dateModified, are covered in our article on freshness signals. The validity dates of the claims, in the text, are the ones that get forgotten; they prevent a model from citing your page for an out-of-date figure, with your name on it.

“Recently”, “currently”, “now”, “this year” cannot be dated. They are to be replaced, systematically, by the period they refer to.

Sources: name them, link them, and know when not to cite

A source cited by name, with the organisation, the document and the year, turns a claim into cross-checkable information. “According to a study” with no name does the opposite: the model detects the formula of authority without being able to verify it, and a third-party source that names the same study will be preferred to yours. Either you name it, or you do not invoke it.

• Name the organisation, the document and the year in the text, not only in a link; the link, where there is one, points to the document itself.
• Prefer the primary source: the report rather than the article that comments on it, the regulatory text rather than its summary. The model cross-checks a source present in its index more easily.
• Separate your data from third-party data. “On our sites” and “according to organisation X” are not cited on the same footing; mixing them in one sentence makes both suspect.
• Do not cite what you have not read. A second-hand source with a distorted figure is the most frequent case of a claim contradicted at cross-checking, because the model has the original source.

The rewritten version looks weaker: it gives up the authority of the anonymous study. It is stronger for the model, because everything it claims is attributed, dated and counted; and it is the only one of the two that a journalist or a directory could repeat, hence the only one that prepares corroboration.

Method: say how you know

Describing the method means answering three questions in one sentence: what the figure covers, over what period, and how it was obtained (measurement, self-reporting, calculation). This sentence goes right after the figure or in a “Method” block at the bottom of the page, and makes possible what the model is looking for: comparing your figure with others while knowing whether they measure the same thing.

A “Method” block at the bottom of the page is useful when several figures share the same origin, provided each figure refers to it and the block is in the page's HTML. For a buying guide, this block is what distinguishes expert content from a compilation.

How a claim is judged: what we know, what we assume

Three things are publicly documented. Generative engines that rely on web search select sources before writing, and their vendors say they favour sources judged reliable and recent, without detailing the calculation. Google publishes quality rater guidelines that stress experience, expertise, authoritativeness and trustworthiness, and the presence of sources and dates; they concern the index on which AI Overviews and AI Mode rely. Finally, models write by comparing several passages: a claim contradicted by the other sources is generally discarded, or cited along with the contradiction.

The rest is observation. On our client engagements, when we classify a page's passages by their number of markers, the cited passages are concentrated among the most marked. The diagram below illustrates this reading; it is not a measurement of the algorithm, it is a prioritisation tool.

Table: scroll horizontally.

The errors that discredit an entire page

An unverifiable claim costs only that claim. Some errors contaminate the page: they signal to the model, and to the third parties who might corroborate you, that the site's figures are not kept up.

• An update date refreshed automatically while the figures have not changed: the “2026” page quotes 2023 rates.
• The year in the footer taken as a validity date, for want of a date in the text.
• A cited source that, on reading, says something else: rounded figure, changed scope, shifted year.
• Two contradictory figures on two pages of the same site: the model retains the inconsistency, not one of the two.
• Round numbers everywhere (“100 customers”, “50%”, “10 years”), which read as orders of magnitude.
• A testimonial with no name, no job title, no date: impossible to cross-check, therefore to cite.

Making content verifiable does not make it longer; it makes it denser, often shorter, because every formula of authority is replaced by a fact. The formats that carry these facts best are described in the previous article; their organisation around an intent, with a pillar page and its satellite pages, in the next one.

What to remember