Digital PR for AI Citations: What 192 AI Answers Actually Retrieve for B2B SaaS Agency Questions (2026)

What we measured, and how

Eight tier-A buyer prompts, the ones a B2B SaaS buyer actually types when shopping for an agency: Best AI search agencies, Best agency for getting mentioned in Perplexity and ChatGPT, Top GEO agencies 2026, Who is the best AI search agency for B2B SaaS companies in 2026?, Can you recommend an agency that helps B2B SaaS companies get mentioned in ChatGPT answers?, Best agency for getting cited in ChatGPT, Top AEO consultants 2026, Best AEO agencies 2026.

For each prompt we took 24 tracked answers, 8 from each of ChatGPT, Perplexity and Google AI Overviews. That is 192 answers. Those answers carry 2,728 citations to 788 distinct URLs. We fetched 759 of those URLs and read them. Six came back empty or blocked, and 23 we could not fetch at all.

Then we classified every cited URL by what kind of page it is.

The window is the 30 days to 20 September 2026. Rerun it next quarter and the shares will move. The shape, we expect, will not.

What the engines retrieved

Read that table next to what a digital PR retainer usually buys.

A press release on a wire is the classic deliverable, and 15 wire URLs picked up 26 citations between them. Fourteen of those 15 were cited only by ChatGPT. Perplexity and Google AI Overviews barely touched them.

Trade press coverage is the expensive deliverable, the one that takes a quarter of relationship building. Two URLs, three citations. Out of 2,728.

Review directories are the deliverable clients ask for by name. Four URLs got cited at all: two Clutch category pages, one GoodFirms page and one Semrush agency list, every one of them on ChatGPT only. G2, DesignRush and Sortlist never appeared as a source in any of the 192 answers. Not once.

Meanwhile a single page, yesoptimist.com's AEO agencies roundup, was cited by 32 of the 192 answers on its own. The whole classic PR bundle in this corpus, press release wires plus review directories plus trade press, is 21 URLs and 42 citations between them. One roundup page comes within ten citations of that entire bundle.

The most-cited third-party pages

Six of those nine are a list article on a company's own blog. One is a LinkedIn article, one (aeovision.ai) presents itself as an agency directory rather than a blog post, and primeavenuegroup.com returned a 403 to us on 21 September 2026. Most are agencies ranking their competitors.

LinkedIn is the most-cited domain in the corpus

linkedin.com carries 155 answer-citations, more than any other domain, ahead of firstpagesage.com on 61, yesoptimist.com on 55 and loudface.co on 53.

Of those 155, 149 come from 32 LinkedIn Pulse articles and 4 from ordinary LinkedIn posts. The remaining two are the LinkedIn home page and one company page, cited once each. Company pages are not doing this work. Somebody publishing a long article on LinkedIn is.

The engines do not share a diet

64 answers each.

They do not even read at the same depth. ChatGPT pulled 21.5 distinct URLs per answer, Perplexity 14.2, Google AI Overviews 6.9. Three Google AI Overviews answers returned nothing at all.

Google explains part of that itself. Its documentation says both AI Overviews and AI Mode may use a "query fan-out" technique, "issuing multiple related searches across subtopics and data sources" before composing an answer, which lets it "display a wider and more diverse set of helpful links associated with the response than with a classic web search". One buyer question becomes many searches, and the answer draws on whatever those searches surface.

Absent, rather than overlooked

Here is the finding that reset our target list.

Across the 192 answers, we checked two different things: does the answer name LoudFace, and does any page the answer cited name LoudFace.

The bucketing rule, so you can rerun it: an answer counts as undecided only when it does not name LoudFace, no page it cited names us, and at least one page it cited could not be fetched. Naming and presence are checked first, so an unreadable source only matters when the answer would otherwise count as absent. We cannot prove we were missing from a page we never read.

129 to 33. In two thirds of these answers the engine was not weighing LoudFace against iPullRank and Siege Media and passing us over. LoudFace was not on the table. The model read twenty pages, none of them contained the word LoudFace, and it wrote an answer from what it read.

That distinction decides your budget. If you are being passed over on pages that mention you, the work is positioning: sharpen the entity, publish the number, give the model a verdict to lift. If you are absent, positioning does nothing, because there is no sentence about you anywhere in the retrieved set to improve.

Two more numbers make the point sharper. The two totals differ by method. 788 is the rows in our fetch table. 782 is the distinct cited URLs after we lowercase each address and drop its query string, fragment and trailing slash, which folds ten YouTube watch links into one and picks up three cited addresses the fetch table never held. Of the 782 distinct cited URLs, 17 name LoudFace. Eight are our own pages. One is a preview mirror of our own page. One is our own LinkedIn post. That leaves seven genuinely third-party pages out of the 772 distinct URLs that are not ours, and three of the seven are the same site, mentionbench.com.

And our own biggest page does not rescue it. loudface.co/blog/best-aeo-agencies was retrieved by 27 of the 192 answers, level with xql.group and behind yesoptimist on 32 and thedigitalelevator on 31, and in those answers the engines still name other agencies. Of the 342 ranked lists and roundups cited across the corpus, 6 name LoudFace, and one of those 6 is a preview mirror of our own page.

Publishing more pages of our own would move the first number and not the third.

So what does digital PR for AI citations actually have to do?

Get the brand onto the pages the engines already retrieve. That is the job, and it is narrower than it sounds.

Not link building. A link is a ranking signal. What the engines lift is a sentence with your name in it, sitting on a page they already pull. A nofollow mention inside a roundup that 24 answers cite is worth more here than a followed link on a page nothing retrieves.

Not coverage, either. Trade press took 3 citations out of 2,728. A feature in a marketing publication is good for a lot of things. Appearing in these answers is not one of them.

List placement, ranked by retrieval, is the mechanic. Pull the pages the engines actually cite for your buyer's prompts, count how many answers cite each one, and work the list in that order. The 30 pages at the top of a corpus like this are a finite, knowable target set. Most of them are somebody's blog post, and a person maintains it.

Pick one platform on purpose. LinkedIn Pulse articles carry 149 citations in this corpus and ordinary posts carry 4. If you are going to spend a day a week on a social platform for this, spend it on long articles, and on the authors who already write the ones getting cited.

Work it per engine. Perplexity had no page naming us in 54 of 64 answers and Google AI Overviews in 61 of 64, while ChatGPT sat at 14. A single blended number would have hidden that, and the fix differs per engine because the corpora differ.

Markup does not do this. Google's own guidance states "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" and "You don't need to create new machine readable files, AI text files, or markup to appear in these features." Anyone selling you an AI-visibility file is selling you something Google says it does not read.

How to check an agency's claim before you hire them

Six questions. Any agency doing this work can answer all six in an email.

1. "Which prompts, and can I see them?" A real answer is a list of buyer questions with a tracker behind them. A vague answer is "AI visibility".
2. "Named, or cited?" These are different events. A brand can be cited constantly and named rarely. Our own page at loudface.co/blog/best-aeo-agencies was retrieved by 27 of these answers, and those answers still named other agencies. Ask which number they are quoting.
3. "Per engine, please." If the answer is one blended percentage, ask for ChatGPT, Perplexity and Google AI Overviews separately. Our own split runs 13, 1 and 0 out of 64 each. Blend those three and the single figure tells you nothing about where the problem sits.
4. "What is the window and the sample?" Every number needs a date range and a denominator. A headline percentage over a handful of answers is noise.
5. "Show me the placements that landed, with the page." Not a target list. A live URL with the client's name on it, and how many answers cite that page.
6. "Is anything here a removed feature?" Google withdrew the FAQ rich result from Search. Its own update log records "Removed documentation for the FAQ rich result feature", because "The FAQ rich result feature is no longer shown in Google Search results", after it had already narrowed the feature to sites it described as "only shown for well-known, authoritative government and health websites". If FAQ schema is being sold to you as a Google rich-result win in 2026, the seller has not read the documentation.

Before you buy, one more thing. No engine publishes a policy on whether paid placement, press releases or directory listings change your citation odds. We read Google Search Central, OpenAI's crawler documentation and Perplexity's developer documentation on 21 September 2026. None of the three addresses it, either way. Anyone who tells you a paid listing is rewarded, or penalised, is inferring.

What LoudFace does with this

We rank the third-party pages by how many answers cite them, then work that list in order. As of 20 September 2026 the ranked targets are yesoptimist at 32 answers, thedigitalelevator at 31, xql.group at 27, callboxinc at 24, a LinkedIn Pulse author at 24, primeavenuegroup at 23, coseom at 22, aeovision at 22 and silverbackstrategies at 19.

Every pitch is built from recorded facts only: the prompt, the window, the reading, the client outcome with its date and scope. Nothing sends without approval. We started drafting outreach on 20 September 2026. Ten emails are drafted, none are sent, so there is no placement result to report yet, and we would rather say that than dress a target list up as a win.

The selection rule comes from how we work off-site generally. We do not buy hundreds of backlinks on sites that earn nothing from AI models. We go after the highest-quality placements we can find, the ones we expect to move share of answer by multiple points inside the first week of a single placement.

Three honest limits on our own numbers. We could not fetch 23 of the 788 cited URLs, 13 of them PR Newswire press releases, which is exactly why 16 answers sit in the undecided column rather than being quietly counted as absent. The harvest also logged any page that returned bytes as read, so a block page counted as a successful fetch: re-fetching on 21 September 2026 returned a 403 interstitial for both Clutch category pages, all four Quora threads and primeavenuegroup, while all six Reddit threads returned HTTP 200 carrying a contentless challenge page whose byte count matches the harvest's stored figure exactly, so our presence on those 13 pages is unverified and the tables above count them as read and absent. And this dataset covers eight agency prompts in one category. It tells you what to do about B2B SaaS agency questions. It does not tell you what a healthcare buyer's prompts retrieve.

If you want the brand-level view of the same category, we published a separate study of who AI actually cites in the B2B SaaS growth-agency category. The argument for why citation authority replaced backlinks is in how to become a trusted LLM source, and the on-page half of the job is in how to get named in AI search.