Which sources AI assistants actually cite for local queries

I would rather start there than give you a tidy pie chart built on somebody's guess. The tidy pie chart is what most of this category sells.

The useful version of this question is not "what does the industry say." It is "what gets cited when I ask about my own category, in my own city, repeatedly." That you can find out this afternoon, for free, and the answer will be about your market rather than a national average.

What does Google actually confirm about its own citations?

Two things, both from its published guidance for site owners.

First, AI Overviews and AI Mode are built on core Search ranking. They retrieve from the Search index and use query fan-out, which Google describes as a set of concurrent related queries generated by the model. So on Google surfaces, the pool of citable sources is the pool of pages that can already rank. Not a different index. The same one.

Second, on local specifically, Google states that generative responses can include information about local businesses and that a Business Profile helps your services appear "in both AI responses and other Google Search results."

That second sentence is the most concrete local AI guidance Google has published, and it points at an asset you already own for free. Which is also why listing your services and products on the profile properly stopped being housekeeping and became a visibility task.

What do the large citation studies actually show?

Two things: user-generated content dominates, and the numbers move fast enough that any of them older than a quarter is worthless.

Similarweb analysed roughly 600,000 citation events from ChatGPT in the US across January and February 2026. Wikipedia at 13.15% and Reddit at 11.97% together accounted for more than 25% of all citations. The Wall Street Journal, the New York Times and Bloomberg did not appear in the top twenty at all.

Semrush, working from a different sample of roughly 150,000 AI citations, found Reddit referenced in 40.1% of AI answers, ahead of Wikipedia at 26.3% and YouTube at 23.5%.

Both are vendor research and both publishers sell in this category. More important than that, look at what the two figures describe. Similarweb counted citation events and Reddit was about a ninth of them. Semrush counted answers containing a Reddit reference and found four in ten. Those are different denominators over different samples in different windows, and combining them produces nothing. One is a share of links, the other is a share of answers. Quote whichever you use, next to what it measures.

Now the part that matters more than either number. Semrush measured ChatGPT citing Reddit on roughly 60% of queries aligned with AI Overviews in early August 2025, and roughly 10% by mid-September 2025. A six-fold swing in six weeks, from one publisher using one method. Any citation-share figure you are shown has that shelf life sitting under it.

What is a citation-source package worth?

Price it against the shelf life. The arithmetic is short.

The quote. An agency offers "placement in the sources AI cites" at $1,200 a month on a twelve month term. That is $14,400.

The claimed basis. Usually the Semrush figure above, presented as though 40.1% of AI answers referencing Reddit means a Reddit post is worth four tenths of an answer to you.

The shelf life. The same publisher watched that share move from roughly 60% to roughly 10% across six weeks. Divide a year by six weeks and you get about 8.7 six-week windows (derived). Your $14,400 buys about $1,655 per window (derived), and inside one observed window the underlying share moved by a factor of six.

The comparison. Ten prompts in your customers' language, three runs each, is thirty answers. Log every cited source. That is roughly half an hour and it tells you what actually gets cited in your category and your city, this month, which is the only version of this number that describes your business.

The honest limit on my own arithmetic. The 8.7 windows figure treats a single observed collapse as a repeating cycle, and nobody has shown that it is one. Use it to price uncertainty, not to forecast. The defensible conclusion is narrower and still decisive: a twelve month contract priced off a share with a six week half-life is a bet, and it should be priced as one.

Which review platforms show up, and how confident can you be?

Less confident than the vendor headline suggests, and the vendor says so themselves.

Whitespark ran 153 queries across 17 local business categories in 9 US cities, none of them cities the researchers were located in, and documented the review sources attached to first-page Bing Places listings. Facebook came out as the most common review source, on roughly 1.5 times as many businesses as Yelp, with Yelp second and more than double the occurrences of third-placed TripAdvisor. Category variance was large: TripAdvisor dominated bakeries and sushi, Porch dominated garage door repair.

Three caveats travel with that, and the authors supply two of them.

It is a proxy. They measured Bing Places, not ChatGPT, on the theory that ChatGPT's local grounding runs through Bing. Whether that theory still holds is genuinely contested, and Whitespark's own conclusion is that there is no firm industry consensus. Both Whitespark and Greg Sterling report not seeing Business Profile data inside ChatGPT.

The source contradicts itself on one detail. The methodology says 17 categories; the findings say Facebook led in 10 of 18. The authors never reconcile the two, so do not present the category count as precise.

And the authors' epistemic warning is the best line in the study. They asked ChatGPT to explain a map it had generated and it denied showing a map at all. AI should not be treated as an authority on its own sources or outputs.

The practical hedge is cheap: claim Bing Places, keep it accurate, stop there. An afternoon, no downside.

Does Reddit get cited because Google and OpenAI pay for it?

This is the causal story everyone reaches for, and it does not hold up cleanly.

The facts are real. Google signed a deal with Reddit worth roughly $60 million a year in February 2024 for AI training rights and structured content access. OpenAI signed a separate arrangement worth roughly $70 million a year in May 2024.

The inference is the problem. Reddit's visibility surge began before those deals fully landed, and the 60% to 10% collapse six weeks apart is not what a contract looks like. Contracts do not renegotiate themselves in six weeks. Retrieval tuning does move that fast. The licensing explains why the content is legally available, not why it is being selected on any given Tuesday.

So do not tell a client, and do not accept from a vendor, that Reddit is licensed and therefore Reddit posts will get them cited. The correct sentence is that Reddit currently gets cited a lot, the share is unstable, and no purchased posting program has been shown to convert that share into your citation.

There is a harder reason to stay away from manufactured Reddit presence. Google names "seeking inauthentic mentions" as ineffective and spam-policed, a burned account in your own city's subreddit is a permanent loss in the exact community you sell into, and in regulated categories it edges toward the federal rule on fake reviews and testimonials. That is not a growth tactic with a compliance footnote. It is legal exposure with a growth story attached.

What can you observe without buying a tool?

Run the prompts yourself and write down who gets named and what gets linked. Ten prompts, three runs each, thirty minutes.

Do that in a few categories and patterns show up quickly. In local queries you will typically see a mix: pages that already rank for the plain search, review platforms, the business's own site, mapping and directory data, local news, and small niche publications that have no business outranking anyone but do.

Record the citations, not just whether you appeared. The citation list is the work list. If the same neighborhood blog gets pulled into four different answers about your category, you now know a place worth being mentioned.

Then check the other side of the pipe. Your server logs will tell you which AI crawlers actually fetched your pages, by IP range rather than by a user agent string anything can type. Sampling tells you what came out. Logs tell you what went in.

Why do small local publications turn up so often?

Because they say specific things that nothing else on the web says.

A neighborhood newsletter writing up three new taco shops on a single street is producing information that exists nowhere else in that combination. A chamber roundup, a school fundraiser sponsor list, a hobby forum thread comparing two shops by name. These are thin, unglamorous pages, and they are dense with exactly the entity relationships a retrieval system needs.

Whitespark's practitioner testing of AI Mode local answers in May 2026 points the same way: star ratings, review volumes, pricing and hours are pulled from Business Profiles or comparable sources, and reviews and unstructured citations, meaning blogs, news and community pages, are the heaviest inputs. That is vendor observation rather than a controlled test, so hold it loosely. It agrees with what Google says about the Business Profile, which is the strongest thing you can say for it.

Big directories, by contrast, say the same thing about everyone. That is not a citation advantage.

Reviews sit in the same bucket of specific, unrepeatable text, which is why reviews function as an input to AI answers rather than only as social proof.

The uncomfortable corollary: that mechanism runs in both directions. A coordinated wave of hostile reviews becomes distinctive text about you too, which is one more reason a review bomb needs handling immediately rather than waiting it out.

What to do this week

1. Write ten prompts a customer would actually type, in full sentences, and run each three times. 2. Log every source cited in every answer, not just whether you appeared. 3. Sort that list by how often each source shows up. The top five are your target publications. 4. Fill in your Business Profile services completely, using the words customers use, not internal names. 5. Claim Bing Places if you have not. It is an afternoon and the local grounding question is unsettled. 6. Pick one local publication from your list and find a legitimate reason to be in it. An event, a hire, a seasonal piece. Real, not purchased.

Before acting on any citation-share figure a vendor shows you, check what is in the denominator, which is most of what comparing AI visibility tools honestly comes down to.

Want to know who is getting cited in your market?

Send your business name and city to eric@seod.com and I will run a set of customer-style prompts for your category and send you back the list of sources that got cited, including the ones that are not you. No charge, no pitch, and you keep the list whether or not we ever talk again.

I do this because that list is the single most actionable thing in this whole subject, and it costs me twenty minutes.

Otherwise, the rest of the AI search library is here.