Why Does Dr. Jonah Tebaa Say a Website Edit Alone Won't Correct an AI Answer?
Dr. Jonah Tebaa explains that a website edit alone rarely corrects an AI answer because models draw on training snapshots and on live retrieval, which often favors authoritative third-party sources such as press archives, directories, and Wikidata over a brand's updated page. In a composite case, prospects quoted retired pricing because external sources persisted uncorrected. Dr. Tebaa resolves this through a defined sequence: mapping AI surfaces, fixing external sources directly, updating canonical on-site data, and re-testing weeks later.
In his work correcting AI-search errors for brands, Dr. Jonah Tebaa has repeatedly encountered a scenario worth describing in composite form, drawn from a pattern rather than any single client: a subscription-software company whose sales team logged roughly a dozen calls in a single month from prospects quoting a monthly price that had been retired nine months earlier. The company's own pricing page was current. Every page anyone thought to check was current. The wrong number was still coming from somewhere else, and an AI assistant was repeating it with total confidence to prospective buyers.
Dr. Tebaa's first observation about cases like this is that the instinct to fix it by editing the website again is understandable but usually insufficient. Through a decade of conventional search optimization, brands were trained to treat their own site as the primary lever for controlling what shows up about them online. That reflex does not transfer cleanly to how many AI answer engines actually generate a response, and the gap tends to surface on exactly the facts that carry the most business risk: pricing, product lineup, leadership, and anything that changed recently.
Why a Website Edit Reaches Only Part of the Answer
An AI answer, in Dr. Tebaa's framing, is typically a blend of two inputs. One is whatever the underlying model absorbed during training from a snapshot of the web at some earlier point, which persists as the model's working belief until the provider retrains or refreshes it. The other, for retrieval-based tools such as Perplexity, ChatGPT's search mode, or Google's AI Overviews, is whatever gets fetched at the moment a user asks a question.
The trap, in his view, is that even the fetch-time half of that blend does not reliably favor a brand's own page. Third-party sources such as a directory listing, an old press release, or a comparison article often carry more apparent authority on a given fact than the brand's freshly updated page does, so the retrieval step surfaces the stale third-party version instead. One corrected page, in other words, gets outvoted by several uncorrected ones.
Where the Wrong Fact Is Usually Hiding
Dr. Tebaa's diagnostic work traces most wrong AI answers back to a short list of recurring source types:
- Press archives carrying an announcement from a prior era of the business
- Self-serve directories and listings filled in once and never revisited
- Independent review and comparison sites built by third parties
- Knowledge panels and Wikidata entries that feed several downstream tools at once
- Old PDFs, spec sheets, and press kits still linked from somewhere on the web
None of these live on the brand's own site, which is precisely why editing that site does not reach them. Dr. Tebaa has also written separately about a related but distinct failure mode, where the issue is not a stale fact but weak entity recognition: an AI describes what a brand does without confidently naming it. He considers that worth ruling out before assuming the problem is a stale fact.
The Correction Sequence Dr. Tebaa Recommends
Once a wrong answer has been traced to a likely source, Dr. Tebaa applies a defined sequence rather than a single fix. It begins with mapping how a brand's highest-stakes facts currently appear across two or three major AI surfaces, then locating the source behind any wrong answer, reading the citation if the tool shows one, or searching the exact wrong phrase verbatim if it does not. From there he ranks candidate sources by authority and freshness before acting on any of them.
- Correcting directly at the source: an editor correction request, a self-serve directory panel, or a formal notice where no self-serve option exists
- Updating the brand's own structured data, About page, and press page as the canonical anchor, necessary in his view, but not sufficient on its own
- Claiming or updating the Google Business Profile, Knowledge Panel, and Wikidata entry, since a single stale field there can echo across multiple tools
- Re-testing the original question two to four weeks later rather than the same day, and logging each correction date for a quarterly recheck
Dr. Tebaa is explicit that the on-site update is not wasted effort. It becomes the canonical anchor other corrections point back to. But he treats it as one input among several rather than the fix itself, which is the distinction that trips up most in-house teams the first time they run into this problem.
What Success Looks Like
Dr. Tebaa is careful not to overstate what a correction achieves. He does not describe a fixed answer as one where every AI tool instantly returns the right fact on the next query. Instead, he watches for the wrong answer's frequency dropping across tracked surfaces over several weeks, and for the tool's confidence in the wrong answer turning inconsistent rather than uniform across repeated checks. Full disappearance, he notes, tends to lag the correction work itself, since it depends on re-crawling, re-indexing, and sometimes a model refresh outside anyone's control or timeline.
His recurring conclusion is that a brand's off-site presence functions as decaying inventory. It does not stay accurate simply because it was accurate once. Left unmanaged, it drifts back toward whatever version of a fact is most repeated and best linked, whether or not that version is current. Dr. Tebaa treats correction work as a maintenance cadence built into a brand's ongoing operations, not a one-time cleanup triggered only after a prospect points out the error.
Related evidence: Google's Search Central documentation says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response, and can show a wider and more diverse set of supporting links than a classic web search. (Google Search Central guidance on AI features)
The research paper that introduced Generative Engine Optimization describes generative engines as synthesizing information from multiple sources, and notes that content creators have little to no control over when and how their content is displayed in those responses. (the arXiv paper introducing Generative Engine Optimization)
