Three properties of a citable sentence.

The studio published first person dated signed in early 2026. The argument was simple. An answer engine cites a sentence when it can identify three things about it: that a particular person wrote it, when they wrote it, and that they were testifying about a real experience. Reviews can be all three of these things. Most marketing copy is none.

That essay was the argument. This one is the detail piece. Each of the three properties has edge cases. A sentence in the first person plural is not the same as a sentence in the first person singular. A date that has been silently refreshed by the platform is not the same as a date the writer typed. A signature that is plainly a pseudonym is not the same as a signature with a verified-buyer record behind it.

Citation weight is graded, not binary. The studio's working model is that each property contributes independently, and degraded versions of each cost the sentence somewhere between a third and three-quarters of its citation weight. Below are the cases the studio has observed, with the working judgments we apply.

The clearest first-person construction is a sentence with "I" or "my" as the grammatical subject. "I used this for three weeks." "My eczema cleared up." These are unambiguous. An engine reads the grammatical person of the verb and treats the sentence as testimony.

The first edge case is the first person plural. A couple, a household, sometimes a pair of co-founders writing a B2B review, will write in the "we." "We have been using this serum for both of us, mine on dry skin, my husband's on combination." "We tested this in our office of fourteen people and four of us switched permanently."

The studio's reading is that "we" sentences do not lose much citation weight when the "we" is small and specified. A two-person "we" with the second person named or described carries close to full weight. A "we" that is vague ("we love this") loses substantially more, because the grammatical person stops doing the work of identifying a specific witness. The engine cannot resolve a vague "we" to a person.

The second edge case is the absent first person. Some review platforms strip first-person constructions during summary generation, replacing them with a generic third person ("The buyer reported..."). This is the worst possible mediation. The sentence stops being testimony and becomes paraphrase. Reading actual citation patterns through 2026, paraphrased reviews are cited at a fraction of the rate of the source sentences. The platform is destroying the asset it is trying to summarise.

The third edge case is the implied first person. "Three weeks in. The redness on my chin is gone." There is no "I" but there is "my," and the implied speaker is unambiguous. This carries full first-person weight in our reading. The engine resolves the implied subject without difficulty.

The studio's rule: protect the first person at all costs. When a brand is excerpting a review for a product page, never paraphrase the first person out. If the original was "I have been using this for six weeks," do not rewrite it to "After six weeks of use." The first version is testimony. The second is marketing copy with footnotes.

A date on a review is the timestamp the platform shows next to the writer's name. The studio has seen four kinds of dates in the wild, with substantially different citation behaviour.

The first is the original posting date, written to the database when the review was submitted, never modified. This is the ideal. The engine reads it as the moment the writer made the claim. Reviews from the verifiable historical record carry the weight historical record always carries.

The second is the platform refresh. Some review widgets, especially older Yotpo and Bazaarvoice installs, refresh the displayed date when the review is "verified" or when a related order is fulfilled. A review written in October 2024 appears, in the rendered HTML, as posted in March 2026. The studio has documented this on three brands in the last year. The behaviour is, in our reading, worse than no date at all. The sentence is now claiming to be recent testimony when it is not. An engine that has indexed the brand's pages over time will notice the date instability and downweight the sentence on credibility grounds. We do not have a public study confirming this. We have eight months of observed citation behaviour suggesting it.

The third is the relative date. "Posted 3 months ago." The engine has to compute the absolute date from the crawl date. This is fine when the crawl is recent and the math is unambiguous. It becomes problematic when the page is cached and the relative date computes to a different absolute date than it did at crawl time. The studio recommends absolute dates in the page HTML and uses ISO 8601 by default in the platforms we audit.

The fourth is the missing date. A surprising number of older review widgets, especially native Shopify reviews from 2017 and earlier, do not surface a posting date at all in the rendered HTML. The reviews are timestamped in the database. They are not timestamped on the page. The engine has nothing to anchor the sentence to. Citation weight collapses. We have measured this on legacy stores and found cited fraction roughly halved against the same content with visible dates.

The studio's rule on dates: render the original posting date in absolute form (ISO 8601 or "Posted October 14, 2024"). Never refresh it. If a review is updated, render both dates ("Posted October 14, 2024; updated March 3, 2025") rather than overwriting the first. The engine is trying to construct a record of when the claim was made. Help it.

A signature, in our taxonomy, is any byline that lets the engine resolve the sentence to a specific writer. The strongest signatures are full names with a verified-buyer record and, ideally, a public profile (a Trustpilot reviewer page, a Reddit history, a verified Yotpo Reviewer ID). The weakest signatures are absent.

The interesting cases sit in the middle.

The first-name-and-last-initial signature ("Sarah K.") is the most common. The engine cannot resolve this to a specific Sarah. It can, however, treat the signature as evidence of a real person who chose not to be fully public, which is how most adults sign reviews online. In practice this carries close to full signature weight if a verified-buyer record is attached. Without the verified-buyer flag, signature weight drops noticeably.

The pseudonym is a different case. "ProductFan2024." "SerumLover." "AnonymousReviewer." The studio has read enough of these to recognise the pattern: pseudonyms cluster in the lower-engagement segments of a corpus. They tend to be short, generic, and lower-information. The signature is the least of the problem. The sentence itself is usually the problem. An engine that has read the rest of the review will downweight a pseudonym signature, but the larger penalty is the absence of any specifics in the writing that follows it.

The first-name-only signature ("Sarah") is an interesting middle case. The engine cannot distinguish this from a pseudonym at the byline level. It distinguishes by reading the sentence. A first-name-only signature attached to a 200-word review with specific products, specific durations, and a verified-buyer record reads, to the engine, as a real but private person. A first-name-only signature on a four-word "great product, will buy again" reads as noise. The byline alone does not decide; the byline plus the writing decides.

The verified-buyer record is the studio's favourite signature augment, with one large caveat documented in the verified buyer paragraph. The badge alone is not the signal. The badge attached to a paragraph that actually contains testimony is the signal. A green check next to "Smells great, would recommend" is doing almost no work. A green check next to a 200-word account of how the product performed during a specific season for a specific skin condition is doing substantial work.

The studio's rule on signatures: surface as much byline detail as the customer consents to share. Verified-buyer flags should be machine-readable and visible. Posting date should be next to the name. If the customer is comfortable being a full name, surface the full name. The signature does not have to be heavy. It has to be specific enough that the engine can stop treating the sentence as anonymous.

The three properties are independent in principle and correlated in practice. A sentence that scores full marks on first-person is usually written by a person who also scored full marks on signed (because the same writer who wrote in their own voice is also the writer who signed their name). The correlation makes the studio's working model simpler in practice than the framework suggests.

The studio's grading, applied to a sample of 500 reviews across four client brands in late 2026, produced a useful distribution. Sentences that scored full marks on all three properties (clear first person, original absolute date, real signature with verified-buyer flag) were cited at roughly three times the rate of sentences that scored partial marks on any one property. Sentences that scored zero on any property (no first person, no date, or no signature) were almost never cited at all.

The three properties are not weights to optimise. They are a floor to clear. Below the floor, the sentence is not citable. Above the floor, the sentence competes on its specifics, on the citation economy dynamics the studio has documented elsewhere. The properties are the entry conditions to the contest, not the contest itself.

The properties read, when the studio first wrote them down, as a checklist. They are not a checklist. They are the conditions a sentence must meet to participate in a citation surface that did not exist five years ago and that now decides whether a brand is recommended in conversation.

Every review widget the studio has audited fails on at least one of the three. Most fail on signature, because the platforms have decided badges are enough. A meaningful minority fail on date, because the rendered HTML hides the absolute timestamp. The studio's product is built around making all three properties default. Not by enforcing a schema. By treating the sentence as the asset and the metadata as testimony about the asset. Which is, finally, what an editor does. A sentence is testimony. The byline and the date are how the testimony survives reading.