First-person, dated, signed.

There are three small properties that, in 2026, decide whether a sentence on the open web survives the journey from publication to citation in an AI answer. The properties are unglamorous. They do not require a content strategy. They do not require a budget. They are, in order, that the sentence be written in the first person, that it carry a date a machine can read, and that it be signed by an identifiable account.

First-person. Dated. Signed. Three words, three properties, the entire shape of what counts as first-person evidence ecommerce content can put on the open web. The wider argument that commerce is now a citation contest, with these three properties as its currency, lives in the citation economy.

A sentence without these three properties is decoration. A sentence with all three is a primary source.

The interesting fact about this list is what falls into each side of it. Almost every paragraph of marketing copy ever written fails on all three counts. Almost every verified-buyer review, written in plain language by a real customer, passes on all three. The gap between marketing copy and a review, in machine terms, is not stylistic. It is evidentiary.

A modern AI search engine ingests text very differently from a 2015-era search engine, and the difference is not a matter of degree but a change in what the engine is even trying to do with the words on the page. The 2015 engine looked at the keywords, the headings, the inbound links, the page title. It was a system for ranking documents by relevance to a query. The 2026 engine looks at the sentences themselves and asks, for each one, two questions. Is this sentence true. Who said it.

The engine cannot verify truth directly, of course. What it can do is weight a sentence by how confidently it can be sourced. A sentence with a named author, a publication date, and a first-person frame is sourceable. A sentence written in the third-person passive voice on a marketing page with no byline and no date is, structurally, anonymous. The engine does not refuse to read the anonymous sentence. It just declines to lean on it.

When the engine produces an answer (a single paragraph, four citations beneath it) the four citations are not the four highest-ranking pages. They are the four pages whose sentences the engine could most defensibly attribute. This is why Reddit threads, dated forum posts, signed reviews, and timestamped news articles keep appearing in answer-engine citations and brand-marketing pages keep not.

The shift in what counts as a useful sentence has happened in less than three years. Most ecommerce teams have not yet sat down with it.

A sentence in the first person says, plainly, I. The author exists. The author is making a claim about their own experience.

Compare two sentences about the same moisturiser.

A marketing sentence. "Customers report dramatic improvements in skin texture within four weeks." Subject is "customers", an aggregate that contains no one in particular. Verb is "report", a neutered version of "said". Object is "dramatic improvements", an adjectival flourish. There is no person in the sentence. There is no date. There is nothing to attribute. An answer engine cannot meaningfully cite this; it is, in source-data terms, hearsay about a non-existent crowd.

A review sentence. "I started this in March and by the end of the second week my skin stopped feeling tight under SPF in the mornings." Subject is "I", a specific human. Verb is "started", a thing that happened on a date the writer is willing to attest to. The claim is bounded, falsifiable, located in a real life. An answer engine can cite this. It will.

The difference is not just rhetorical taste. The first-person sentence is parseable as testimony. The third-person aggregate is parseable as marketing. Language models trained on the open web have learned the distinction over the past three years of ingesting forum posts, signed reviews, dated news articles, and marketing copy side by side, and they weight accordingly.

The implication for an ecommerce brand is awkward. Most of the brand's site is written in a voice that the engine cannot use. The product page, the about page, the blog, the homepage. None of them speak in the first person. The only place on the brand's site where actual first-person language exists, almost always, is the reviews. And the reviews are presented as decoration under the buy button, in a widget the crawler often cannot render.

The asset is on the site. The asset is invisible.

A date on a published sentence does two things. It tells a crawler when a claim was made. It also tells a crawler the claim has not been silently re-dated to look fresh.

Google's Helpful Content signal, folded into core ranking since March 2024, penalises the refreshed-evergreen pattern that took over the SEO playbook between 2019 and 2023, when every brand was told to keep their old posts looking new. The tactic of taking a 2023 post and editing the date to read 2026 is now actively suppressed. The suppression applies in inverse on AI search: a real, original publication date earns trust. A date that visibly matches the time the content was first written is a small but compounding citation signal.

A verified-buyer review carries this property for free, in a way the brand can't manufacture. When the customer hits submit, the platform stamps the record with a timestamp the brand can't go back and edit. That immutability is the whole game, because what the engine is actually checking for isn't just "is there a date" but "has the date been quietly moved", and a review timestamp tracked at the platform layer answers the second question as well as the first.

Marketing copy almost never has any of this. About pages on a typical Shopify store carry no date, anywhere. Product pages don't say when they were last edited (and if they did, they'd be saying "yesterday" most of the time). Blog posts sometimes carry a date, but it's the date of whatever recent edit kept the post looking fresh, not the date of original publication. None of it is useful to a crawler trying to fix a sentence to a moment in time.

What a citation-friendly dated sentence actually looks like, then, is something more specific than "a sentence with a date next to it". It looks like a first-person paragraph at a stable URL, timestamped at the time of writing, on a page the crawler has visited before and seen at the same date, in a corpus where the dates around it form a continuous timeline. Continuity over months is part of the signal. The single date alone isn't enough.

A review system that quietly does all of this (publishes verified-buyer language on indexed pages, preserves the original timestamp, leaves the URL stable) is doing the single highest-leverage thing a brand can do for AI citation, and almost no brand is doing it. The reviews are dated. The page they live on isn't. And the page is usually rendered by JavaScript the crawler may not even run. A closer look at the specific sentences answer engines lift, taken from a working store, sits in what ChatGPT reaches for when it recommends a serum.

A "signature", for our purposes, is anything the crawler can resolve to a real account. Full name not required. A verified-buyer record on a review platform counts. So does a Reddit handle with a posting history, a Trustpilot profile with two or three other reviews on it, a LinkedIn byline. The point is just that the sentence isn't drifting around the page anonymous, attached to no traceable person.

What a signature gives an answer engine is twofold. First, somebody to hold accountable for the claim, if the claim turns out to be wrong. Second, evidence that the person exists outside this single sentence, which is roughly the difference between a witness and a sock puppet. An identifier with a history reads as the first. An identifier without one reads as the second.

So "Anna B., verified buyer" is, in the math of citation, dramatically more valuable than "five-star anonymous review", and not by a small margin. The first carries a checkable identity. The second carries nothing. (The deeper reframing of reviews as language, rather than as starred inventory, sits in a separate piece for anyone who wants to follow that thread.)

It's also why marketing copy, signed by nobody, or signed by "the brand" (which is to say, signed by nobody), occupies a structural floor on the engine's evidence ranking. The brand isn't a witness to its own product; the brand is the entity with the incentive to lie about it. Every modern answer engine knows this and routes around it. The engine reaches past the brand for the witness, which is to say, for the customer.

Here is what the three properties look like in practice, on a real timeline, for a real piece of customer language.

A woman in Manchester orders a sensitive-skin moisturiser in May 2026. She uses it for four weeks. On May 14, she leaves a review on the brand's review platform. The review is two paragraphs long. It says, in the first person, that she has rosacea, that she had been using a different brand that stung on application, that this one did not, and that she has now bought it twice. She signs it with her first name and last initial. The review platform timestamps it, attaches it to her verified-buyer account, and publishes it on the brand's site.

That is all that happens on May 14. The brand's marketing team does not touch the sentence. The customer is not contacted. The widget under the buy button now contains one more review.

Over the next four months, several things happen quietly. The review is indexed by Google. The review is crawled by GPTBot, PerplexityBot, ClaudeBot, and Google-Extended. The sentence "did not sting on application like the [other brand]" is parsed as a first-person, dated, signed claim about a specific product property. The engine notes it.

In October 2026, somebody types into ChatGPT: "what is a good moisturiser for rosacea that does not sting". The engine returns a paragraph that mentions the brand, with four citations beneath it. One of them is the May 14 review. The customer in Manchester has, without knowing it, become a primary source.

A second customer reads the answer, follows the citation, lands on the product page, buys the moisturiser. The brand counts a new sale and does not know why.

The why is on the review page. The why has been on the review page since May. The why is first-person, dated, and signed.

It changes very little about the brand's own writing. The brand should keep writing brand-voice copy on the homepage. The brand should keep the about page. Marketing copy has its job: it sets the register, it sells the moment of intention, it carries the visual identity.

What it changes is what the brand publishes outside the marketing voice. The brand should treat its review corpus as the publication, not the decoration. The reviews should appear on pages a crawler can render. The pages should preserve the timestamp, the verified-buyer signature, and the original first-person voice. The brand's reply, when it writes one, should also be signed and dated, in the brand's name, not "Customer Support Team".

The brand should resist the urge to clean up the first person into something more on-brand. The first person is the asset, to be edited lightly for typos and never paraphrased into marketing.

The brand should resist the urge to fake the signature, which arrives most often as a quiet suggestion from an agency deck that promises a shortcut to the citation page. Anonymous reviews, fake verified-buyer badges, AI-generated testimonials with invented names are not just unethical; they are also detected. The Federal Trade Commission's 2024 rule on fake reviews is the legal floor. The engines have a parallel floor of their own. A signature that does not resolve to a real account is, increasingly, a downgrade signal, not an upgrade one.

The brand should publish the verified, signed, dated reviews on pages with the same care it spends on the homepage. These pages will, over the next two years, do more work than the homepage. They are the brand's primary sources.

The three properties are unfashionable. They were unfashionable when reviews were displayed as stars, and they are unfashionable now that everyone wants to talk about AI strategy. They are unfashionable because they do not let the brand write the sentence. They put the sentence in the customer's mouth and ask the brand to step aside.

That is the discipline. The brand steps aside. The customer says, in the first person, what happened. The platform records the date. The account carries the signature. The engine, six months later, reads the sentence and decides it is the one worth quoting.

The work of the next two years, for any brand that wants to be cited, is to take this discipline seriously and to publish accordingly. First-person. Dated. Signed.

Everything else is decoration.