GEO-045·Issue 04·Authority & Off-site·45 / 132
Authors and Editors: AI Needs to Know Who's Accountable
A byline is not accountability. What belongs on an author page, how to log review and corrections, and when putting a name on the page makes trust worse.
Issue focusAnswer Units: Writing Category and Product Pages for GEO
Two pages sit open in a Tuesday review at a Colorado DTC tent brand selling into the US and Germany. Its own guide runs 3,000 words, signed “Editorial Team”. A rival’s is shorter but signed by a person: fourteen nights in the Wind River Range, two tents, one storm at 12 °F. Ask a chat assistant which tent holds above treeline: the answer paraphrases the second page, never the first.
AI quotes a paragraph only after it can tell who would stand behind it, so a byline, a review trail, and a correction log are the accountability chain both models and buyers read.
Anonymous pages fail on the second pass
Retrieval finds your page; generation assembles the sentence. Both ask who is speaking: a minimum accountability unit needs a name resolving to a person, an organization resolving to a legal entity, and a date. “Editorial Team” and “Admin” resolve to nothing outside the company, so responsibility stays inside, where no reader can reach it.
Google’s quality rater guidelines ask human evaluators to weigh experience, expertise, authoritativeness, and trust — public documentation, current as of April 2026 — not a ranking formula. How much weight a model gives a byline is unpublished, and unmeasurable is not unused: a byline might win one layer of trust; anonymity gives one up.
The counter-example sits in most footers: “Compiled by our operations team, for reference only.” A model reads a company that will not answer for the sentence; a buyer reads an ad.
An author page is three checkable facts
The test: could a stranger check most sentences? A page that fails is decoration.
| What to write | Passes when | Fails when |
|---|---|---|
| Relationship to the category | Seasons, terrain, models tested, credentials | “10+ years in the industry” |
| A trail off your domain | Same identity on LinkedIn, trade press, an expo roster | One avatar, only on your site |
| Interest boundaries | Paid by brands or not, how editorial relates to sales | Silence, which readers fill with “you are selling” |
The middle row is skipped most and matters most: someone who exists only on your domain is, to a model, indistinguishable from an invented person; a name matching two independent sources becomes an entity. Which is why named authors are a supply line: planned ahead, not found on deadline.
Mark the block up: Person, sameAs to outside profiles, datePublished, dateModified. That markup only helps if extraction works — paragraph boundaries are part of the same job, as HTML semantics for extraction sets out. Bylines also pair with answer units: the unit says what was said, the byline says who is answerable.
Review, correction, and change history are states
“Professionally reviewed” carries no information. Review takes three roles — writer, fact-checker, compliance reviewer — even when one person holds two.
| Check | Ticked when |
|---|---|
| Numbers have a source and a date | Report ID, official document, or your test record, dated |
| Units and market match | Ounces and grams, °F and °C, US and EU voltage never mixed |
| Claims stay inside the line | No medical efficacy, no absolutes, no unprovable “first” |
| One claim, one version | Product page, guide, FAQ, and support script agree |
One line at the foot of the page does the trace: who wrote it, who checked facts, when it was last reviewed. Not polish — internal process turned into a state an outsider can read.
“Updated April 2026” proves a file was touched, not that it was fixed. A change log takes three columns: date, change, reason. Two honest updates a year beat fifty date bumps; a silent edit leaves citations quoting the old version, and the mismatch reads as inconsistency.
Keep the correction entry obvious — a form at the page bottom with one stated order: verify, correct, then log. A correction log is a trust asset, not a PR liability.
A five-step playbook that starts with the archive
| Step | Action | Owner | Passes when |
|---|---|---|---|
| 1 Inventory | Sort guides into named author, editorial team, no byline; flag cited pages | Content lead | Every page has an owner |
| 2 Author pages | Track record, outside trail, interest boundaries, Person markup | Content and front-end | Each person verifiable off-site |
| 3 Checklist | Turn the five checks into a template; tick or write “not checked” | Fact-check | Five pages re-checked, no blank rows |
| 4 Trace | Three footer lines plus a three-column change log | Editor | Any page explains what changed |
| 5 Retest | Same prompts; do answers carry the author’s name? | Growth | A before and after you can explain |
Step one is not new writing. Lay the archive out and many pages will have no accountable owner; that belongs in the planning meeting, not a content queue.
Example: from “Editorial Team” to one checkable person
Example — hypothetical brand, no real company, no outcome figures. A German storefront carries twelve guides signed “Editorial Team”. Week one: move the byline to whoever tested the products, and publish an author page with three facts: which models, under what conditions, whether any brand pays. Week two: add the three footer lines and a change log for six months of price and regulation edits; where no date exists, write “no earlier record” rather than inventing one. Week three: rerun the prompts.
No revenue number belongs here; orders are not the first signal. What you watch is the author’s name appearing in answers, or the answer quoting the conditional test rather than an adjective — sampled with the method in an AI visibility baseline, not one screenshot.
Where this breaks
Marketplaces. Listings have no author field; a byline in A+ content reads as odd. Work the brand store and reviews.
OEM and white-label. An NDA may forbid naming the client; write “the brand’s engineering team” with a lab you can name.
Small teams without an expert. The founder can sign, but describe what they did: units, factories, failures. An invented “chief scientist” is worse than anonymity once checked.
Pages with a version, not an author. Size charts, shipping terms, and return flows need an effective date.
Bulk-generated copy nobody read. A byline there is fabrication. Either a person reviews it, or the page becomes specs and data.
Regulated categories. Medical, children’s, financial: a byline without a license number is a liability.
The authority hub collects this line. This month, do one thing: pick the ten guides closest to being cited, add three footer lines to each — checkable author identity, checker and date, first change-log entry — and publish one author page. Any guide whose checker cannot be named is downgraded to “reference notes”.