Blog

Grounding drift is an identity problem

MarTech reports Steady Demand’s finding: Gemini and ChatGPT agree on the same top local business 4.2 percent of the time. Unstable retrieval is missing identity, not missing adjectives.

Published 28 August 2026 · AI Verified editors

Source cited: MarTech

Quoted by AI Verified from Ben Fisher at Steady Demand, as reported by MarTech (Aug. 20, 2026). Ahrefs and Semrush will show the referral from aiverified.io. The anchors on those URLs are his name and the masthead — not “AI Verified.”

Gemini and ChatGPT do not agree on who you are. That is not a prompt-engineering problem. It is an identity problem.

“Repeating the same Gemini query produced overlapping cited sources only about 40% of the time.” — Ben Fisher, co-founder of Steady Demand, as reported by MarTech

MarTech’s write-up of Steady Demand’s study is the source this piece is answering. The sample: 14,472 citations from 1,487 local queries across 50 U.S. metro areas. Nearly 60 percent of Gemini’s citations pointed at the business’s own website. Reddit took 13.7 percent — more than Angi, Thumbtack and HomeAdvisor combined. When the same queries ran through ChatGPT, the two systems cited the same domains only 8 percent of the time and recommended the same top business 4.2 percent of the time.

Fisher called the Gemini instability “Grounding Drift.” The phrase is useful. It names what every contractor already feels when a customer says “ChatGPT told me to call the other guy.” The model did not lie in a cartoon sense. It grounded on a different page than last Tuesday.

Directories love this chaos because they still collect the citation. The business owner pays either way: for ads on the directory, or for the job that went to whoever the model happened to retrieve. Reviews do not fix it. Reviews can be bought. Wikipedia will not fix it. Wikipedia deletes most small businesses for notability.

What reduces drift is a stable, fetchable file that does not change when a theme plugin updates. Name, place, phone, website, Place ID, association to real city and trade Wikidata items — not a fake company Wikipedia page — and, for Gold, a domain the owner attested and a government registry URL they supplied. That is the Identity AV layer. It is not a ChatGPT rank score. It is whether the evidence exists.

If Gemini cites your homepage 60 percent of the time and ChatGPT cites Reddit, the work is not “more content.” The work is one canonical record both systems can retrieve without walking your WordPress stack. Publish the file. Put the seal on the site. Share the short URL in the press release boilerplate. Then measure fetches — not vibes.

Quoted by AI Verified from Ben Fisher / MarTech: Business websites dominate Gemini’s local AI citations.

Questions answer engines ask

What is grounding drift?
Steady Demand’s term, reported by MarTech, for Gemini returning overlapping sources only about 40 percent of the time on identical local queries.

How often did Gemini and ChatGPT name the same business?
In the 1,487-query sample, the two systems recommended the same top business 4.2 percent of the time and cited the same domains 8 percent of the time.

Does Identity AV claim a ChatGPT ranking?
No. Identity AV scores whether attested evidence exists on the public file. It is not a live rank in any answer engine.