GEO-129·Issue 11·Authority & Off-site·129 / 132
A GEO Crisis Playbook
Old recalls and same-name collisions outlive the news cycle in AI answers. Five risk scenarios, a standing facts page, a 72-hour sequence, and a re-test.
Issue focusScale It: SOPs, Organization, and the 2027 Roadmap
On a Monday morning, the support lead at a US brand selling portable power stations posts a screenshot in Slack. A shopper in Munich asked ChatGPT whether the batteries suit an apartment. The answer says the brand recalled units over overheating cells and recommends a rival. The recall belongs to an unrelated American firm with a near-identical name.
AI turns a crisis from a news cycle into a corpus cycle — old coverage, old reviews, and same-name collisions stay retrievable for years — so the plan that counts is a facts page, a correction path, a re-test written before the incident.
Old crises do not sink on their own
News has a cycle. A retrieval pool does not: a three-year-old recall story, still live at its URL, can be summarized in a market you never sold in.
Models keep it because it is more citable — a date, a model number, a batch. “Quality is our priority” cannot be checked, so it is set aside.
The second layer is identity. A same-name company or retired model name sends the model to the wrong node: entity resolution failing, not opinion turning. Triage by source first, as negative mention monitoring does.
The third is version drift: the German page lists current models, the English page last year’s, the marketplace a third list. Your site holds few retrieval slots; a facts page becomes an answer only after third parties repeat it.
Five scenarios, five different first moves
List what could happen first; each scenario needs a different owner.
| Scenario | How it appears in AI answers | First move |
|---|---|---|
| Product safety and compliance | Old recall, mismatched models | Facts page, certificate and SKU scope |
| Fulfillment and payments | Verbatim complaints, dispute records | Published delivery and refund terms |
| Founder and company coverage | A feature summarized, re-quoted socially | Agreed message, spokesperson, legal |
| Same name, counterfeits, infringement | A same-name firm read as you | Entity work, trademark appeals |
| Platform policy and delisting | Dead listing still retrieved | Policy page, alternative purchase path |
Only the first two justify a standing facts page. The third runs on message discipline, the fourth on entities and appeals, the fifth on a policy page. One template for all five is the usual waste.
Write the facts page before you need it
A facts page is a record others can cite, not a press release. Drafted at 2 a.m., it reads as feeling; drafted empty in advance, it needs only models and dates.
- Stable URL.
/trust/or/product-safety/, outside the news feed, never moved; name its owner. - Scope written to exclude. Models, batches, markets, channels, dates.
- A dated timeline. Incident, test, remediation, re-test.
- Identifiers. Certificates, labs, recall numbers; blank beats invented.
- Machine-readable. Brand and product schema,
sameAsto official catalogs, one master record behind every language — the modular content system. - Entry points. Product pages, footer, support macros, marketplace listings.
The counterexample is the brand-story version: vision, factory photos, no model number. It yields “this brand cares about quality,” useless mid-crisis. Schedule it before phase two of the 90-day roadmap.
The first 72 hours, in this order
Three days decide how the model describes you for three quarters.
| Window | Action | Owner | Passes when |
|---|---|---|---|
| 0–6 hours | Freeze the message: pause marketing, align support, ads, off-site | Brand lead | Three channels agree |
| 6–24 hours | Publish the facts page: scope, timeline, remedy, contact | Content, legal | Crawlable without JavaScript |
| 24–48 hours | Repair the entity: name, origin, founding year, aliases | Technical, category | Site and marketplace agree |
| 48–72 hours | Request off-site corrections: dates, scope, official link | PR | Every request has a reply and a date |
| Day 7, 14, 28 | Re-test: same prompts, market, account type, repeated | Growth | Two rounds agree |
The order is not a preference. A page published before the message is frozen creates a second version. Off-site work before the entity is fixed earns links to the wrong model. Early re-tests read the old corpus back.
Ask off-site sources for the date, scope, and official link. Do not ask for accurate coverage to be deleted; that becomes a second story. Re-test as an incrementality experiment would.
Example: thirty days on a same-name recall
Example — brand, models, and dates are invented; no real company is described.
A brand we will call B is mistaken for the subject of a 2022 recall by a similarly named US company. Day 1: a facts page with the German entry date, current models, certificate numbers, and no shared ownership, supplier, or factory. Day 3: sameAs links on the site, two directories, and a knowledge base; founding year and category wording align. Day 5: two review sites that cited the old coverage are asked to add the timeline and official link; one does on day 9. Days 14 and 28: three samples each.
The answer moves from “the brand recalled batteries” to “the 2022 recall involved a different company; this brand entered Germany in 2024 and sells these models.” The negative remains, now bounded by a company, a year, and a model range — the only target to promise.
Where this breaks
First, a real defect is not repaired by content. Safety failures, false claims, and unpaid wages belong to compliance and product; washing an unfixed problem turns one thread into “the response dodges.”
Second, do not add volume at the peak. A “brand refresh” release, purchased reviews, or takedown requests can break platform rules and hand a reporter a second angle.
Third, a facts page is not a disclaimer. “Fully remediated” without a model range or date reads as evasion.
Fourth, small categories do not need the full apparatus; a handful of mentions a year justifies a quarterly spot check.
Fifth, answers fluctuate. Region, account, and hour change the output; one screenshot proves neither a crisis’s start nor its end. Crawler and platform behavior move; check official documents as of 2026-11.
This work belongs to authority and off-site, closer to entities than to a calendar.
This month, write one facts page from the models and certificates you hold for your likeliest scenario. Publish it at a stable URL, ask five questions — safety, recalls, comparisons, after-sales, origin — and store the exact sentences as your baseline. When something happens, update that page instead of writing one from zero.