Two stale blog posts made Google's AI Overview deny a new API

Steve McLeod, who founded the player feedback tool Feature Upvote, searched “feature upvote api” in July 2026 and got a Google AI Overview telling him the platform had no REST API, a week after he had launched one. Semrush’s case study traces that answer to two old blog posts on Feature Upvote’s own site, each of which said in passing that the product had no REST API.

The new help documentation existed and carried “Public REST API” as its H1, and Google used the older pages anyway. One reason surfaced during the repair: McLeod’s announcement post from July 24 was not in Google’s index at all, contrary to what he had assumed. On the question Google was being asked, the only evidence in its index was two stale denials written by the company itself.

McLeod had published that announcement across a blog post, an X post and an email newsletter, partly as promotion and partly to hand Google fresh data. Tan Siew Ann, a content writer for B2B software companies who wrote the Semrush piece and knows McLeod, read the newsletter and wrote back on July 27 with three suggestions: correct the outdated pages so they state the platform does have a public REST API, submit the new and updated pages to Google for indexing, and get the API covered in industry publications. The first two were under a day’s work, so those went first.

Within an hour of submitting the pages on July 28, the query “does feature upvote have an api?” returned an AI Overview saying yes. Tan rechecked the next day in three separate browser tabs and in incognito, to rule out a one-off or a personalized result. The press coverage never happened and turned out not to be needed.

The likelier explanation is that the two indexed denials were holding the answer in place, whatever the new documentation said. A page on a company’s own domain that denies a fact about its product is evidence Google can quote, and it keeps working for as long as it stays indexed. The safer assumption is that a new page asserting the fact competes with the old page denying it rather than replacing it, which makes an audit of old posts worth more than another new doc. The fact also has to sit in text that can be quoted back: facts placed only in markup came back in zero of 12 AI answers in one practitioner test.

The correction did not reach every phrasing at once. Per Semrush’s write-up, McLeod found that “does feature upvote have a rest api” returned yes while “does feature upvote’s rest api include this” still returned no, and that it took a few more days before the answer was consistent. So a single query coming back right is not evidence the fix has landed. On September 10 the answer still said yes, but the AI Overview was citing different pages than the announcement post, so the sources behind a correct answer can shift without warning.

Semrush’s piece does not make this distinction, but the indexing request is a Google-only lever, because AI Overviews are generated from pages in Google’s index. Other assistants find pages through their own search indexes or fetch the page live when asked. The edit to the old post is the part of the fix that carries to all of them.

What to do

• Pick the facts that cost you money when an AI answer gets them wrong: features and integrations, pricing, service area, operating hours, refund policy, regulatory compliance, and name, address and phone.
• Write three to five phrasings per fact, run them on the platforms your audience actually uses, and log the date, the answer word for word, and the pages cited. Keep the wording identical on later runs, so a change in the answer is attributable to your fix rather than to your prompt.
• Search your own domain for pages that state the old version of the fact, old blog posts and release notes included, and correct the sentence that gets it wrong.
• Put the fact on a standalone page as a full sentence near the top, in the words people search with (“Yes, Feature Upvote has a public REST API”), and repeat it in an FAQ answer if the page suits one.
• Request indexing for every edited and new URL in Search Console’s URL Inspection tool, and confirm the pages are indexed before concluding the content did not work. McLeod’s announcement had been live for days and was invisible to the answer that contradicted it.
• Rerun the same prompts after the fix and again after any product change, since the pages an AI Overview cites can change while the answer stays right.