Advertising was built on assertion. You said your detergent washed whiter, your bank cared more, your software scaled better, and the game was saying it louder and prettier than the next brand. Humans could not check most of it, and the entire industry was engineered around that inability. I have written those claims. I have shipped campaigns for global brands where the strongest line in the deck was the least checkable one.
That era is ending, not because regulators finally caught up, but because the audience changed. When an AI assistant summarizes your category for a buyer, it does something no consumer ever did at scale: it attempts to verify. It looks for the claim elsewhere. It checks whether anyone independent repeats it, whether your numbers appear in any source you do not control, whether your "award winning" traces to an actual award. Claims that survive that process get repeated by the machine, in the machine's voice, which is the most valuable endorsement in modern marketing. Claims that fail it do not get argued with. They get silently omitted, and omission is the new rebuttal.
So the discipline I now teach clients is not claim writing but claim engineering. Every important assertion about your brand should sit at the center of what I call a Corroboration Web: a deliberate lattice of independent, crawlable, mutually consistent sources that a machine can traverse to confirm the claim without ever trusting you.
Key Findings
- Verification is becoming a default behavior of answer engines, not a premium feature. Assume every public claim will be cross checked against sources you do not control.
- Unverifiable claims are not penalized, they are dropped. The cost of puffery is invisibility in machine mediated answers, which is harder to detect than any penalty.
- A claim's machine value is set by its weakest link: precision of wording, stability across your own properties, and existence of at least one independent attestation.
- Self published corroboration counts for little. The web must include surfaces you demonstrably do not own: registries, press, customer sites, public records, third party databases.
- Specific, dated, falsifiable claims outperform superlatives. Machines can confirm "founded in 2014, offices in Orlando and Dubai" and can do nothing with "world class."
- Brands that restructure their top twenty claims for verifiability in 2026 and 2027 will be quoted as fact by assistants while competitors are paraphrased with hedges, or skipped.
- Run the claims inventory this quarter. Every public factual assertion, one spreadsheet, three columns: is it atomic, is it anchored, who are its witnesses. Kill or rewrite anything that fails all three.
- Apply the Three Witness Rule to your top twenty claims. For each missing witness, assign an owner and a deadline, exactly as you would for a product feature, because that is what it is.
- Rewrite superlatives into falsifiables. Every "leading," "premier," and "world class" becomes a date, a number, or a named proof, or it comes out of the copy.
- Ask your five best clients this month to state the relationship in their own words on surfaces they control. It is a five minute favor and it is worth more than a testimonial page you host yourself.
- Re run your top claims through the major assistants every quarter and record which ones the machines repeat as fact. That repetition rate is your real message penetration, and I would put it on the same dashboard as pipeline, because within two years it will be upstream of it.
The Verification Turn
The first generation of language models repeated whatever they had absorbed, which briefly rewarded whoever published the most confident nonsense. That window is closing fast. Retrieval augmented systems now ground their answers in live sources, agents click through and compare, and model builders are under existential pressure to reduce fabrication. The direction of travel is unmistakable: every serious answer engine is becoming a verification engine, a shift I flagged early in my answer engine optimization guide.
Here is the operational consequence. Your brand's claims are no longer evaluated in the persuasive environment you built for them, the landing page with the testimonial carousel. They are evaluated in a cold, comparative environment where the machine holds your statement in one hand and the rest of the observable web in the other. If those two hands disagree, you lose. If the other hand is empty, you also lose, just more quietly.
I want to be precise about the failure mode, because marketers keep imagining the wrong one. The machine will rarely call you a liar. It will simply construct its answer from claims it could confirm, and yours will not be among them. Your competitor whose duller claim was corroborated in four places becomes the quoted fact. You become background noise. No dashboard will alert you to this. You will just watch pipeline soften and blame creative.
Anatomy of a Verifiable Claim
Not all claims can be engineered for verification, and knowing the difference is half the skill. A verifiable claim has three properties.
It is atomic. One fact, cleanly separable. "We serve 400 clients across 12 countries" is two atomic claims stapled together, and each can be checked or broken independently. Compound claims fail as units even when their parts are true.
It is anchored. It carries the details a checker needs: dates, names, places, units. "Award winning" is unanchored. "Winner of the 2025 X Award for retail campaigns" hands the machine a query it can actually run.
It is attested. Somewhere beyond your control, the same fact exists: an award body's own list, a registry filing, a press mention, a client's site, a conference program. This is the load bearing property, and it is the one marketing departments systematically neglect because it cannot be produced in a content calendar. It has to be earned or requested.
My working standard is what I call the Three Witness Rule: a claim is not machine real until it appears on three surfaces, at least two of which you do not control. One witness is an assertion. Two is a coincidence a cautious system might discount. Three independent appearances form a triangle the machine can stand on. Inside my own firm, no claim goes into a client's core messaging until we can name its three witnesses or a plan to create them.
Weaving the Web
Corroboration webs do not happen by accident, and they are not built by publishing more. They are built by exporting your facts into other people's hands.
Start with the claims inventory. List every factual assertion your brand currently makes in public: about us pages, sales decks, bios, packaging, ads. In my experience the list runs between sixty and two hundred items, and the first audit is always humbling. A third are stale, a tenth are contradicted by your own properties, and most have zero independent witnesses.
Then triage ruthlessly. Pick the twenty claims that actually move buying decisions and machine summaries: what you do, for whom, since when, at what scale, with what proof. These get the engineering budget. The rest get archived or rewritten as opinion, which machines treat differently and more forgivingly than false precision.
For each of the twenty, manufacture witnesses deliberately. Get the client to state the relationship on their own site. Get the award body to publish the winners list. File the registration. Pitch the trade publication on the story that happens to contain the number. Update the industry database. None of this is glamorous, and all of it outlasts any campaign, because a corroborated fact keeps testifying for you in every answer, every night, for years.
Finally, make the web traversable. Machines can only verify what they can crawl and parse, so structured data, stable URLs, and clean crawler access are the plumbing under all of this. I covered the technical layer in my guide to preparing websites for AI crawlers, and I will only add: a corroboration web behind a broken robots file is a library with the doors welded shut.
The Comparative Trap
One dynamic deserves its own warning. Verification is comparative, which means your claims are checked not only against reality but against your rivals' claims about the same reality. When two brands assert leadership of the same category, the machine does not flip a coin. It weighs webs. Whoever has more independent attestation for their version wins the phrasing of the answer, and the phrasing of the answer is the market.
This creates a genuinely new competitive surface. Imagine two agencies of similar size. One spends 2026 producing thirty more portfolio pages. The other spends it getting fifteen clients, three award bodies, and two trade journals to independently state facts about their work. By 2028, when the majority of B2B shortlists begin life as an assistant's answer, the second agency is described in confident declaratives and the first in vague maybes. Same reality, different webs. I know which side of that trade I am taking, and it is the reason verification, not persuasion, is where I now spend my clients' authority budget. Persuasion still closes deals. Corroboration decides who gets into the room.