AI Cites Your Own Website First: What 14,472 Local Citations Reveal

Muhammad Shahid·AI-Powered Digital Marketing Consultant

Google Ads Certified

Meta Blueprint Certified

Google Analytics Certified

Updated August 2026

AI assistants get their local recommendations mostly from the businesses themselves. In new research analysing 14,472 citations across 1,487 local-service queries in 50 metro areas, the business's own website was the source behind 59.9 percent of Gemini's local citations. Reddit and community forums followed at 13.7 percent, general business directories at 10.4 percent, and the combined local-service directories such as Angi and Thumbtack at 10.3 percent. If you have been told your visibility in AI now depends on getting listed somewhere else, the data points the other way.

The dataset is Steady Demand's AI Citation Ledger, collected on 30 July 2026 with follow-up verification on 20 August 2026, and covered by Search Engine Land. It is US metro data across ten local-service categories, so read the exact percentages as a direction rather than a forecast for your suburb. The mechanism it exposes is what matters, and it holds in the accounts I run.

What did the study actually measure?

It measured which sources an AI assistant leans on when someone asks it to recommend a local service business, and how stable those answers are. The researchers ran 1,487 queries across 50 US metros and ten service categories, captured 14,472 citations pointing at 3,611 unique domains, and recorded the 4,410 businesses Gemini named. They then compared those against 9,364 businesses pulled from plain Google search as a control group.

Marie Haynes summarised the practical read on the study page: “60% of AI citations still point to the business's own website. That's a huge lever you still control!” That is the part I would hold onto. Most of the anxiety I hear from owners about AI search is about surfaces they cannot touch. The biggest single source is the one they already own and usually neglect.

Why does your own website carry so much of the weight?

Because the assistant needs something to back up a claim, and your site is the only place that states your services, service area, pricing approach, and process in your own words. Gemini uses retrieval-augmented generation, which means it fetches live pages and cites them rather than reciting memory, the same grounded approach Google describes for AI Mode. When it wants a sentence about what you do, a clear page on your site is the cheapest thing for it to quote.

What that changes about your site

A homepage that says “quality service since 2011” gives a retrieval system nothing to lift. A page that says which suburbs you cover, what a callout costs, how fast you answer, and what you will not take on gives it four extractable facts. In the sites I rebuild, the pages that start getting cited are almost always the ones that answer a question in the first two sentences.

• One page per service, each answering “what is it, what does it cost, how fast” up top
• Service areas written as plain sentences, not a keyword list in the footer
• An FAQ block using the questions customers actually ask on the phone
• LocalBusiness and FAQ schema so the facts are machine-readable, not just visible

Why is Reddit beating Angi and Thumbtack in AI answers?

Reddit accounted for 13.7 percent of Gemini's local citations, more than the entire local-service directory category combined at 10.3 percent. The reason is not that Reddit is a better business directory. It is that a thread where three people argue about which plumber turned up on time reads like evidence, while a directory listing reads like an advertisement. One of those is useful for backing up a recommendation, a split the citation breakdown makes hard to miss.

Joy Hawkins, quoted in the same research, said what a lot of local marketers were thinking: “I was surprised that Facebook wasn't higher on the list. I see that site constantly on Google.” The sources that fill a Google results page and the sources that back an AI answer are not the same set, and assuming they are is how people end up spending on the wrong listings.

I want to be careful here, because this gets misread fast. This is not an argument for going and posting about your own business on Reddit. That gets you removed, and it is the kind of shortcut that ages badly. It is an argument for being the business people mention when someone else asks. Angi, HomeAdvisor, the Better Business Bureau, and Reddit all appeared in every one of the 50 metros studied, so presence in those places is table stakes rather than an advantage.

What is grounding drift, and why should you care?

Grounding drift is the study's term for the fact that the same question does not return the same answer. With zero change in wording, repeated Gemini queries produced overlapping cited domains only about 40 to 46 percent of the time. The follow-up search terms Gemini generated for itself overlapped about 5 percent. The gap held at p < 0.0005 across a 2,000-shuffle permutation test, so this is not noise in a small sample. The published methodology runs that comparison over 5,000 pairwise rounds for Gemini against 4,466 for Google.

Ninety percent against seven percent is the number I would put on the wall. Dustin Stout put the implication plainly in the study: “They don't rank you, they sample you...optimizing for one engine's algorithm is a losing game.” Volatility also varies by trade. Dentists were the most unstable vertical in the data and locksmiths the most stable, which fits what I see: the more crowded and similar the local market, the more the assistant shuffles.

The practical consequence for owners

Stop treating one prompt as a verdict. I have watched an owner see his business named by Gemini, screenshot it, and conclude the AEO work was done. I have watched another owner see a competitor named once and assume he had been penalised. Both were reading a single draw from a shuffled deck. What you want is a share of voice across repeated runs, which is a number that moves slowly and honestly.

Do Gemini and ChatGPT recommend the same businesses?

Almost never. Cited domains overlapped between the two engines just 8 percent of the time, and they named the same top business in only 4.2 percent of comparisons. Auto repair had the highest agreement at 14.9 percent of domains. Personal injury law had the lowest at 0.5 percent, which is close to two entirely separate universes answering the same question, as Search Engine Land's write-up of the study set out.

ChatGPT also moved sharply between the two collection dates, which is worth seeing as a sequence rather than a snapshot:

1. On 30 July 2026, Reddit and community sources made up 41.7 percent of ChatGPT's local citations, with business websites at 15.9 percent.
2. By 20 August 2026, Reddit had dropped to 0 percent and business websites had risen to 42.8 percent.
3. General directories moved from 34.6 percent to 40.8 percent over the same three weeks.

A source category going from 41.7 percent to zero in three weeks is the clearest warning in the whole dataset. Anyone who built a visibility strategy on ChatGPT's Reddit weighting in July had it deleted by August. Krystal Taing's note in the study applies to all of it: “Different engines are pulling from different sources...one prompt is a snapshot.”

Do star ratings decide who gets recommended?

This one surprised me. The correlation between a business's rating and how often it was cited came out at 0.013, which is effectively no relationship at all. Businesses the AI recommended averaged 4.75 stars, slightly below the 4.84 average of the plain Google baseline. Ninety-seven percent of AI-recommended businesses sat at 4.0 stars or higher, against 99.6 percent of the baseline. Worth remembering that Google still prohibits review gating and bulk solicitation, so there is no shortcut on that side either.

Review volume told a similar story. AI-recommended businesses averaged 919 reviews against 846 for the baseline, but the medians ran the other way at 295 against 322. Fourteen percent of AI-recommended businesses had fewer than 50 reviews. So a thin review profile is not automatically disqualifying, and a strong one is not a ticket in.

I am not going to tell an owner to stop collecting reviews, because reviews still decide whether the person who reads your name actually calls you. What this data kills is the idea that pushing from 4.6 to 4.8 will buy you AI citations. It buys you conversions. Those are different jobs, and I cover the split in map pack versus AI search.

What should you actually do in the next 30 days?

Five things, in this order. None of them require a new tool subscription.

1. Rewrite your top three service pages answer-first. Put the plain answer to the customer question in the first two sentences. This is the 59.9 percent lever, and it is the one you own outright.
2. Add LocalBusiness and FAQ schema so your hours, service area, and answers are readable as data, not just as design. Google documents the local ranking factors it uses publicly, and the same facts feed the assistants.
3. Run a ten-repeat test. Pick five questions a customer would ask, run each ten times in both Gemini and ChatGPT, and record how many of the fifty runs name you. That percentage is your baseline.
4. Check your details agree everywhere. Google Business Profile, Yelp, Facebook, and your site should state the same name, hours, and phone number. Disagreement gives a retrieval system a reason to pick someone else.
5. Keep your profile active. Posts, photos, and updates keep the freshest version of your facts in circulation. Andrew Shotland's figure in the same study is a useful reality check on all of this: “Only about 12% of AI results were clickable on desktop, and only about 28% on mobile across 15.2M local searches.”

The bigger pattern

Two years of local SEO advice trained owners to think about placement. Where do I sit, who is above me, how do I move up one spot. This data says the assistants are not doing placement at all. They are assembling an answer from whatever they can find and defend at that moment, and your website is the thing they find first 60 percent of the time.

That is genuinely good news for a small operator, and I do not say that often. The one asset that carries the most weight is the one asset nobody can take from you. It also sets a floor on how much a volatile answer should worry you: if you disappeared from a Gemini answer this week, the base rate says you would have disappeared anyway. Fix the inputs, measure a distribution, and stop reading single results. If you want that measured properly for your own business rather than guessed at, that is what the audit covers, and the ongoing work sits under AEO and GEO alongside the local SEO side. For the volume side of the same story, the AI local visibility data is worth reading next.

The questions below are the ones owners have sent me since this study landed.