Corrections
Everything we have got wrong, kept here permanently with the date we found out. We do not silently edit pages. A site that grades other people's evidence has no business hiding its own mistakes, and the cheapest way to prove that is to publish them before anyone catches them.
2 corrections logged
-
001 2026-07-10 corrected before publication /evidence/#answer-instability What we said
SparkToro found that AI assistants return an identical brand list less than 0.1% of the time.
What is actually true
SparkToro found that the same brand list repeats less than 1 time in 100 (under 1%). The 1-in-1,000 figure, about 0.1%, applies to getting the same list in the same ORDER. The two are routinely conflated.
How we found out
The error entered our own research notes from a secondary summary. We caught it while fetching the SparkToro post directly to verify the figure before writing about it. We had already repeated the wrong number once in our working notes.
The lesson
This is the exact failure this site exists to catch, and we committed it on day one, on the single claim we were most confident about. A real statistic drifted by a factor of ten in the space of one retelling. Nobody lied. Somebody paraphrased.
-
002 2026-07-10 corrected before publication /methodology/ What we said
The first draft of the grading scale had five grades and no way to mark a claim that a good study disproves.
What is actually true
It graded the popular claim 'you can track your brand's ranking position in AI answers' as VERIFIED, because we had verified the SparkToro study that demolishes it. That is backwards. A sixth grade, REFUTED, now covers it, and the scale is explicit that a grade describes the claim as it circulates, never the finding that settles it.
How we found out
We found it while building the page that renders the ledger. Two claims came out reading as though we endorsed the thing we were disproving.
The lesson
A grading scale that cannot express 'this popular claim is false' is not a grading scale, it is a citation list.
Why the first entries are our own
Both corrections above were made before this site had a single reader. That is not a coincidence, and it is not humility theatre. The failure this site exists to catch, a real number quietly drifting as it is retold, is not something other people do because they are careless. It happened to us on the claim we were most confident about, within hours of starting.
If we had not gone back to the primary source, we would have published it. That is the entire argument for going back to the primary source.
Corrections as a credibility signal in a noisy field
The GEO and AI-search marketing fields launched with extraordinarily low evidentiary standards. Vendor research papers are cited as gospel without verification. Statistics circulate for years with no locatable primary source. Case studies are published with no control group, no methodology, and no contact information for independent verification. Bloggers cite other bloggers citing press releases citing internal vendor claims, and nobody checks the first link in the chain. In this environment, publishing corrections is not humility—it is a structural competitive advantage. It signals that you have actually read the sources you cite, and that you care whether they say what you claim they say. Most GEO and AEO content does neither.
The discipline of corrections forces exactness upstream: if every claim might be checked and published wrong, writers become more careful about which claims to make. Vendors often do not publish corrections because their business model does not reward precision; it rewards confidence and market share. A tool that publishes "we now support 40% visibility uplift" in a marketing deck builds brand momentum. A tool that later publishes "that 40% was a benchmark, not a ChatGPT measurement" loses that momentum. A corrections page is profitable only if you do not need the momentum more than you need the trust of readers who check your work.
This page exists to signal that geooptimised.com is the site where a reader can be confident that a number has been checked, and if it is wrong, the wrongness is on this page in permanent form. That is not a minor advantage in a field where most readers have never seen a retraction and most vendors have never published one.
How corrections work in the AI-search marketing landscape
Most GEO and AEO vendors and content sites do not maintain a corrections section at all. Of those that do, most treat corrections as public relations crises to be minimised: a small notice at the top of an article, unpinned from the homepage, possibly deleted after a year. geooptimised.com takes the opposite approach. Every correction, no matter how small, lives here permanently. The date it was discovered is recorded. The lesson drawn from the mistake is explained. Nothing is silently edited, and no article narrative is changed without an entry on this page explaining what changed and why. This approach makes corrections more costly internally—the organisation cannot quietly improve its track record without acknowledging every improvement. It also makes corrections more valuable externally: readers know that a finding here is not revised on the basis of new evidence alone, but revised with a public record of the revision, dating the moment the site changed its mind.
The broader field does not follow this model. Comparisons sites (especially tool marketplaces) frequently alter rankings and descriptions to respond to vendor complaints, product changes or new partnerships, all silently. Academic papers can be revised after publication (arXiv versions are numbered and dated), but the popular retellings of those papers—the blog posts, guides and vendor claims—are usually edited without the original publisher's knowledge or permission. The moment a number is cited wrong, the error propagates downstream and becomes very difficult to correct after the fact. Publishing corrections is the practice that could slow that propagation. It is rare because it is expensive: it requires someone to check, someone to fact-check, and someone to publish the error on the public record rather than pretend it never happened.