What AI answers actually cite
when you ask for a marketing agency

01 — The short answer

More than half the time, it cites nothing

Of 140 reads, 76 produced an AI Overview with zero clickable sources and 64 produced one with at least one. That is the first thing anyone selling you “getting cited in AI” should tell you and usually does not: on the queries where an agency is chosen, the most common outcome is an answer that names no one.

When the block does cite, it is generous — 5.7 distinct domains on average, between 2 and 10, drawn from 57 different domains over the window. So this is not a surface that cites one winner. It is a surface that either cites a handful of pages or cites nobody, and which of the two you get depends more on the shape of the query than on who is best.

And the composition of that handful is the part that reorders a content plan: 47 of the 64 citing reads included at least one video, forum or profile platform — YouTube, Reddit, LinkedIn — rather than only agency-owned pages.

02 — The record

Who actually gets cited

One read per query per day, 30 days, 13 queries, same API, geolocated to Miami. The queries are the ones a business owner or a marketer types when they are shopping for this kind of work: “generative engine optimization agency”, “GEO agency”, “AI automation agency”, “marketing agency miami”, “digital marketing agency miami”, “social media agency miami”, “SEO agency miami”, “local seo services”, “best generative engine optimization agencies”, “pharma marketing agency AI search”, “agencia de marketing con IA”, “agencias de marketing con IA en LATAM”.

Source: August's own measurement panel, field geoMentions, surface “Google AI Overview”. Percentages are the share of the 64 reads that cited at least one source, not of all 140 reads. No client metrics appear here: these are market queries and no client site is named. Readings taken on other surfaces are not shown on this page, because readings from different sources do not compare.

03 — What it means

The lever is not your services page

Read the table by what each cited domain is rather than by its name, and a pattern shows up that no services page can act on. The most cited single domain is a video platform. A forum is fourth. A profile network is twelfth. These are places where someone else publishes about you, or where you publish in a format most agencies do not produce at all.

The agency sites that do appear — and several appear in a quarter to a half of the citing reads — did not get there with a better “Our Services” page. They got there with a specific published answer to the exact question in the query: what GEO is, how it is measured, how long it takes, what it costs. One page per question, not one page per service. That is the whole difference, and it is reproducible.

The directories are the surprise in the other direction. Clutch appeared in 15 of 64 citing reads and no other directory appeared at all. In a classic result page for these same queries, directories and listicles take half the first page. Inside the answer they take about a quarter of the citations, in one brand. A paid listing is not a strategy for this surface.

Which leaves the 54% where nothing is cited, and that is not dead space. When the block names no sources, the attention it does not spend on a citation lands on what sits underneath it — the organic results and, on anything with a city in it, the map pack. We measured the same behaviour on local searches and wrote it up separately: AI Overviews cite no one on local searches. The conclusion is the same in both: being citable and being findable are two jobs, and you need both because you do not get to choose which kind of answer your buyer gets.

04 — Our own number

August appears in none of the 140

Zero. Not named, not cited, in 140 reads over 30 days on the 13 queries that describe exactly what we sell. The same panel that produced the table above produces that line every morning, and it goes into the client-facing record with the rest.

We publish it for a reason that is practical rather than noble. An agency that shows you only the readings that went well is showing you a shortlist, and you have no way to tell the difference from the outside. So here is the test, and it costs you one email: ask whoever is pitching you AI-answer visibility for their own measurements — their domain, their queries, their method, their dates, with the bad days in. If what comes back is a client case study, you learned something.

What we are doing about our own zero is the same thing described above, run on ourselves and logged: one page per real question instead of one page per service, mentions where other people publish, and a daily read so that the week it changes, we can date it. The log is public: August on August.

05 — What this does not say

One method, one city, 30 days

140 reads on 13 queries, geolocated to one metro area, over 30 days, with one method. That is enough to call the pattern consistent and not enough to call it a rule about Google. Agency queries in another city, in another language, or for another vertical can behave differently, and the only way to know is to read them.

Two specific limits worth stating plainly. First, “cited” here means a domain shown as a source inside the Overview block — not a mention inside the prose without a link, which this method does not capture. Second, the window starts on September 18, 2026 and not earlier even though the panel holds older reads: the reads before that date came from a different capture with a different query set, and splicing them onto these would be comparing two sources. The longest honest window is the longest window measured one way.

There is also no number from any other surface on this page. We read ChatGPT and web search too, on overlapping queries, and those readings live in their own records. Putting them next to these would produce a chart that looks like evidence and is not.

How answer engines choose which sources to cite · Which agency should you hire to get cited in ChatGPT

06 — What to do with it

Four moves, in this order

First, measure your own queries. Pick the eight or ten a buyer would actually type to find a firm like yours. Read the answer once a day for two weeks and record two things: did it cite anyone, and who. That is your baseline, it costs nothing, and without it nobody can ever prove to you that they moved something.

Second, convert your services page into questions. One page per question your buyer actually asks — what this is, how it is measured, how long it takes, what it costs — because every agency site in the table above got in with one of those, not with a capabilities deck.

Third, go where the citations come from. On this record that means video and the places other people publish: a channel that answers the same questions out loud, and being mentioned in writing by someone who is not you. It is slower than a landing page and it is what the surface rewards.

Fourth, keep the classic work running for the 54% of answers that cite nobody. On those reads there is no citation to win, and the click goes to the result underneath. An AI-visibility plan that lets the ordinary ranking rot is betting the whole budget on the coin landing one way.

07 — Questions

What people ask about this

08 — Keep reading

The rest of the log

Pieces from the same measurement series, held to the same rule: every figure with its date and its method.