What AI Engines Cite for GTM Agency Questions (October 2026 Study)

What do AI engines cite when buyers ask for a GTM agency?

Mostly agencies writing about agencies. On October 22, 2026 we asked three AI models with web search turned on the same 30 questions B2B buyers ask about GTM, outbound, cold email and Clay agencies. The 90 answers carried 504 citations to 386 distinct pages on 244 domains. Agency websites supplied 70% of those citations, and 241 citations (48% of everything cited) pointed at "best agencies" style list pages published by an agency, often one that ranks itself.

Software vendor blogs were the second biggest source at 16%, led by Salesforge. Directories, partner listings and review sites were 6%. Reddit, G2, YouTube and Wikipedia were not cited once. Clutch was cited once.

Which domains were cited most?

Counted as the number of answers that cited the domain at least once, out of 90. A domain counts once per answer no matter how many of its links appear. Domains cited in 6 or more answers:

Astra GTM (astragtm.io) was cited in 3 answers: once by Perplexity for "fractional GTM engineer for hire" and twice by Claude, for "outbound agency vs in-house SDR" and "should I hire a GTM engineer or an agency".

Which individual pages were cited most?

Every page in this table is a ranked list, a pricing explainer, or a comparison published by an agency. None is an independent review.

Which software vendors, directories and media sites were cited?

Every non-agency domain cited in at least 2 of the 90 answers, counted the same way as the domain table above. LinkedIn and Google Maps pins are left out. Domains cited in a single answer, and the exact pages cited, are in the CSV linked at the end.

Which types of sources do AI engines cite?

We sorted every cited domain into one source type by hand, from what the site sells. "Looks like a ranked list" means the URL path contains words like best, top, alternatives, agencies, companies or vs. It is a pattern check, not a page read.

• Agency sites dominate because agencies publish the most pages that match the question word for word: "best GTM engineering agencies", "best Clay agencies", "cold email agency pricing".
• Software vendors publish the same list format to rank and convert. Salesforge alone was cited in 13 answers, through both its blog and its agency directory.
• Community sources were absent. Reddit, G2, YouTube, Wikipedia and Quora had zero citations across all 90 answers.

Do the engines behave the same way?

No. The three engines differed far more from each other than the question topics did.

• Perplexity Sonar cited sources in every answer, about 10 per answer, and supplied 300 of the 504 citations.
• Claude Haiku 5.5 cited sources in every answer, about 5 per answer.
• GPT-4.1 mini often answered with Google Maps place links for each agency instead of sources. We excluded those pins from every count on this page because a map pin is not a source.

How does this compare with our September 22 baseline?

We ran an earlier version of the same probe on September 22, 2026. It used different models and an overlapping but different question set, so treat this comparison as directional only.

September counts were raw citation counts per question, not answers, so we compare which domains show up rather than the numbers. Of the 13 domains cited in 6 or more October answers, 5 were also among September's 30 most-cited domains: Understory, Salesforge, LinkedIn, ColdIQ and DevCommX. Salesforge was near the top both times.

What does this mean for B2B buyers choosing an agency?

• Treat an AI shortlist as a reading list, not a verdict. Nearly half of what these engines cited was an agency's own list of agencies.
• Check who wrote each cited list and whether it says so. The better lists disclose that the publisher is on the list and state their criteria.
• Go to primary sources for the facts that matter: the agency's own pages for minimums and scope, partner directories like Clay Experts for partner tier, and review sites for named reviewers.
• Ask what you keep when the engagement ends: the workflows, data, copy and sending setup, and whether they live in your stack or the agency's.
• Ask for named results you can verify and talk to a reference. Our own are on our case study pages. For example, Clipper Defense went from zero outbound to an average of 23 ICP-fit meetings a month.

How was the study run, and what are its limits?

• Engines: openai/gpt-4.1-mini and anthropic/claude-haiku-5.5 through the OpenRouter API with its web search plugin, and Perplexity's sonar model through the Perplexity API. These are API models, not the consumer ChatGPT, Claude or Perplexity apps, which use different models and settings.
• Prompt: a short system prompt asking for current web sources, concrete company names and cited URLs, then the buyer question as written. Temperature 0.2 for the two OpenRouter models and the default for Perplexity. Max 1,200 output tokens.
• Single run: each question was asked once per engine on the evening of October 22, 2026 (US Central). AI answers vary from run to run, so small counts can move a lot.
• Small model tiers: these are fast, low-cost models. Larger models may cite differently.
• Questions: 30 questions we wrote to match how buyers search for GTM, outbound and Clay help. A different question set would produce different sources.
• Missing question text: our merged answer file saved the question only for Perplexity's 30 answers. For the other 60 (Claude and GPT) the question field was empty. We re-joined each of those answers to its question from the raw per-call files, which are named by question number and wording, so the per-question data in the CSV is complete. We flag it because it is where an error would most likely hide.
• Counting: a citation is a URL in the engine's citation list or a link in the answer text. Each URL counts once per answer, and domains count once per answer in the domain tables. Google Maps pins are excluded.
• Classification: source types were assigned by hand per domain from what each site sells. Fifteen citations we could not place sit under "Unclear or off-topic".
• What it is not: citation counts are not traffic, rankings or quality scores. Being cited often does not make a list accurate.

How do I cite or embed this study?

Quote, chart or republish any number here with a link back to this page. The CSV is free to reuse with the same attribution. If you spot an error in the data, email rees@astragtm.io and we will fix it and note the change on this page.

Bayba, R. (2026, October 22). What AI engines cite for GTM agency questions (October 2026 study). Astra GTM. https://astragtm.io/playbooks/ai-answer-citations-gtm-agencies-2026

<a href="https://astragtm.io/playbooks/ai-answer-citations-gtm-agencies-2026">What AI engines cite for GTM agency questions (Astra GTM, October 2026)</a>

<blockquote cite="https://astragtm.io/playbooks/ai-answer-citations-gtm-agencies-2026">
  <p>In 90 AI answers to B2B agency questions, agency websites supplied 70% of citations, and 48% of everything cited was a "best agencies" list published by an agency.</p>
  <footer>Source: <a href="https://astragtm.io/playbooks/ai-answer-citations-gtm-agencies-2026">Astra GTM, What AI engines cite for GTM agency questions (October 2026)</a></footer>
</blockquote>

1. 01Raw citation log, October 22, 2026 (CSV, 504 rows) One row per cited URL per answer: engine, question, URL, domain, source type, ranked-list flag.
2. 02How to measure LLM visibility for B2B GTM The measurement framework this probe follows.
3. 03Why AI answer engines don't mention your brand