Every local business has now been told that AI assistants are recommending competitors. Almost nobody has been told what those assistants are actually reading when they do it, because until August 2026 nobody had measured it at scale.
Ben Fisher of Steady Demand ran 1,487 local business queries across 50 US metros and logged every source the assistants cited: 14,472 citations in total. It is the largest published look at local AI answers to date, and the results contradict most of what the local SEO industry has been selling, including a few things we have said ourselves.
The same question, a different answer
Start with the finding that changes how any of this should be measured. Ask an AI assistant the exact same local question twice and it names the same top business about 7% of the time. Google’s local pack, asked the same way, returns the same business about 90% of the time.
How often repeating an identical local question returns the same top business, against roughly 90% for Google’s local pack
Source: Steady Demand, “What AI Actually Cites,” 14,472 citations across 1,487 local business queries in 50 US metros, August 2026 — steadydemand.com/ai-citation-ledger/
Change the wording slightly and it gets looser still: different phrasings of the same question shared cited sources only 40% of the time. A local AI answer is not a ranking that holds from one day to the next. It is a draw from a pool of businesses the assistant considers plausible.
If a provider shows you a single AI answer as proof of where you stand, they are showing you one coin flip and calling it a position.
That has a practical consequence for anyone buying AI visibility work. A one-off screenshot proves nothing, in either direction. The only honest measurement is the same prompt run repeatedly and reported as a rate: you appeared in 18 of 25 runs, your competitor in 22. Anything else is noise presented as a number.
Your own website is the biggest local citation source
The second finding cuts against a decade of local SEO habit. When AI assistants cited a source for a local recommendation, 60% of those citations pointed at the business’s own website. That is more than directories, forums and review platforms combined.
Of local AI citations point to the business’s own website, more than directories, forums and review platforms combined
Source: Steady Demand, August 2026, 14,472 local AI citations — steadydemand.com/ai-citation-ledger/
Directories still matter for the Google local pack, and reviews still matter everywhere. But in AI answers specifically, the single largest source is the business’s own site, which raises the value of the unglamorous work: whether AI crawlers can reach the pages, whether the important content sits in the raw HTML rather than only in the rendered page, and whether the headings say plainly what the business does and where.
One source does punch above its weight, and it is not a directory. On Gemini, Reddit alone accounted for 13.7% of local citations, more than every local service directory combined at 10.3%.
The assistants barely agree with each other
There is no such thing as a single AI visibility score, and this data is why.
| Comparison | Agreement | What it means |
|---|---|---|
| Gemini and ChatGPT citing the same domains | 8% | The two assistants are reading almost entirely different sources |
| Gemini and ChatGPT picking the same top business | 4.2% | Winning on one tells you nothing about the other |
| Brands appearing in all three major models | 11% | Cross-platform presence is rare and has to be built deliberately |
| Brands appearing in only one model | 77% | Most businesses are visible in exactly one place and do not know which |
Platform agreement: Steady Demand, August 2026. Model coverage: Fractl, 4,320 responses and more than 8,500 brand references across GPT-4o, Gemini 2.5 Flash and Claude Sonnet 4.6, August 2026.
A blended cross-platform score averages four disagreeing systems into one number that describes none of them. Worse, it hides the useful information: which platform you are already winning, and which one your competitor owns.
Ranking still feeds the answer
None of this means classic search work stopped mattering. When AI Overviews launched in France, SE Ranking mapped 52,564 answers across 100,000 keywords and found that 60.46% of the URLs cited in those answers already ranked in the organic top 10, and 76.33% ranked in the top 50.
Of URLs cited inside AI Overviews already ranked in the organic top 10 for the query
Source: SE Ranking, 100,000 French keywords across 20 niches, 52,564 answers with citation data, August 2026 — seranking.com/blog/ai-overviews-france-study/
Organic position is an input to AI citation, not a competitor to it. The businesses that do well in AI answers are usually the ones that were already visible, with the technical and reputation work to back it up.
What this means for a local business
One prompt is not a measurement
Ask any provider showing you AI visibility numbers how many times they ran each prompt, on which platforms, and on what dates. If the answer is once, the number they gave you is a coin flip. Repeated runs reported as an appearance rate is the only defensible method, and it is not expensive to do.
Your website carries more of this than you were told
With 60% of local AI citations pointing at the business’s own site, the technical basics stop being hygiene and start being the product: crawler access for the right bots, critical content and schema in the raw HTML, headings that state the service and the city in the words customers use.
Per-platform, or it is not a report
Two assistants agreeing on the top local business 4.2% of the time means a single score is meaningless. Ask for the breakdown. Then fix the platform where the gap is widest, rather than spreading effort evenly across four.
Instability cuts both ways
The 7% repeatability is usually read as bad news. It is also the opening. A competitor who appears today is not entrenched, and the consideration set is being redrawn on every query. In a category where nobody has been named consistently, being the business the assistant learns first is still available.
We sell third-party citation and placement work, so it is worth saying plainly: the 60% own-site finding cuts against a directory-first view of local AI visibility, and that includes some of how this industry, ourselves included, has framed citation building.
What the data supports is a split. Directories and reviews still carry the Google local pack and still feed the assistants’ sense of who a business is. But for AI answers specifically, the business’s own site is doing the heaviest lifting, and third-party work earns its keep through topical fit in the right category rather than through volume. We would rather tell you that than sell against it.
Free Report · 5 Business Days
See how often AI actually names your business
We run 15 prompts five times each across ChatGPT, Perplexity, Gemini and Google AI Overview, report your appearance rate and recommendation rate per platform, and show the exact run count, dates and model versions behind every number.
Related research
- Ghost citations: why being cited by AI is not the same as being recommended
- How local businesses actually get recommended in AI search
- What the largest AI search study of 2026 shows about local business visibility
- What the data says about AI search in 2026
- Your email list is also your AI recommendation audience
- Four-platform local visibility benchmark report 2026
Local AI citation data: Steady Demand (Ben Fisher), “What AI Actually Cites,” August 2026. 14,472 AI citations across 1,487 local business queries in 50 US metros — steadydemand.com/ai-citation-ledger/.
Model coverage data: Fractl, August 2026. 4,320 responses and more than 8,500 brand references across GPT-4o, Gemini 2.5 Flash and Claude Sonnet 4.6 — searchengineland.com/ai-visibility-index-brands-vanishing-from-ai-search-485057.
AI Overview citation and ranking data: SE Ranking, August 2026. 100,000 French keywords across 20 niches, 52,564 answers with citation data — seranking.com/blog/ai-overviews-france-study/.