When you ask an AI assistant who to hire to fix your positioning, it does not read vendor homepages and pick the best one. It assembles the answer from third-party comparison roundups and listicles, plus a thin layer of community threads and earned media. The firm that gets named is the one already cited across the pages the model trusts, not the one with the sharpest website. Read every AI vendor recommendation as a summary of who gets written about, not a verdict on who is best.
- AI does not read vendor sites to answer "who should I hire." It summarizes the pages that already rank and get cited.
- Comparison roundups and listicles are the dominant source these answers pull from.
- Community threads and a little earned media round out the rest. Vendor homepages barely register.
- The field is fragmented. No single firm dominates, and most firms a founder would name never appear.
- You earn the recommendation by being cited across the pages the model already trusts. That is relevancy engineering.
A founder finishes a board call where the numbers were fine and the story was not. Positioning came up. That night they open ChatGPT and type a version of what a lot of founders now type: who should we hire to fix this. The answer comes back fast, confident, and specific. Three or four firms, a sentence of praise each, a tidy comparison.
It reads like judgment. It is closer to a summary.
The model did not evaluate those firms. It cannot call their references or inspect their work. What it can do is read the pages already written about the category and repeat the names that appear most across the ones it trusts. The confidence is real. The basis for it is thinner than it looks.
We wanted to know exactly what that basis was. So we watched one question get answered, over and over, for two weeks.
How AI builds a "who should I hire" answer
Ask a person for a consultant referral and they draw on memory, reputation, and who they trust. Ask a model and it does something narrower. It retrieves the pages it can reach that match the question, weighs the ones that look authoritative, and composes an answer from what those pages already say.
That means the answer is only ever as good as the sources feeding it. For a factual question, that is often fine. For "who is the best firm to hire," it quietly swaps the question you asked for a different one: which firms are named most across the pages I trust. Those are not the same question. A founder reads the first. The machine answered the second.
The gap between them is where this gets useful to understand.
How this was measured
One model, ChatGPT, sampled daily for roughly two weeks through August 4, 2026, US-based. We tracked which pages and domains it cited when answering five vendor-selection prompts, and how the named firms were spread across them:
- "Who can help my startup own a category"
- "Category design agency vs positioning consultant"
- "Best strategy consultants for Series A and Series B startups"
- "Best positioning consultants for early-stage startups"
- "Alternatives to hiring a PR agency for a startup"
Limitations, stated plainly: one model, five prompts, two weeks, one country. This is a directional read of how these answers get built, not a census of the market. Treat the pattern as real and the precise proportions as approximate.
What two weeks of watching one model showed
The answers come from roundups, not homepages
Almost every citation was a third-party comparison page. "Best B2B positioning firms," "top GTM consulting firms," and X-versus-Y comparison articles were the dominant format the model pulled from. The firms' own websites rarely got cited at all. The homepage a firm spent months polishing was not where the machine looked.
Community and earned media fill the rest
A thin second layer rounded the answers out: a handful of Reddit threads, one Forbes council piece, a few LinkedIn company pages. Enough to shade the recommendation, not enough to carry it. The center of gravity stayed on the comparison roundups.
The winners are a fragmented long tail
The firms that surfaced were mostly small specialists sitting on the right comparison pages, not the names a founder would recognize. Only one prompt behaved differently. The pure "own a category" question was held by a few established category-design firms. Every other question was wide open.
No brand dominates the category
Across the five prompts, no single firm owned the answer. Share of voice was spread thin and shifted with the sources the model happened to pull that day. Most firms a founder might name from memory did not appear at all. In this category, the machine has no default. It has a reading list.
How to read a firm that AI recommends
Once you know the answer is a summary of coverage, you can read it correctly. A recommendation tells you a firm is written about. It does not tell you the firm is good, or a fit for your stage, or capable of the specific thing you need. Those are separate questions, and the model did not answer them.
So use the list for what it is. It is a fast way to find candidates and a poor way to rank them. Ask the assistant where each name came from, then read the original source. A named client or a piece of published work is worth more than a slot on a generic list. From there, the diligence is yours: inspect the work, call the references, ask the firm to explain its method. Order in the answer is not ranking, and confidence in the answer is not proof.
Read this way, the AI recommendation stops being a verdict and becomes what it always was. A starting point.
Being the recommended answer is a relevancy problem
Here is the part that matters if you are the firm, not the founder. On these questions you do not become the recommended answer by having the best homepage. The machine barely reads homepages. You become the answer by being named inside the roundups, comparisons, and threads the model already trusts.
That is a different job than building a nice website, and most firms are still doing the wrong one. Being the cited source across the pages a category is built on is relevancy engineering. It is how a firm earns a position that the market repeats and the machine can retrieve, instead of a claim that lives only on a page nobody cites.
The founders reading these AI answers deserve to know how the sausage is made. The firms hoping to appear in them should stop optimizing the one surface the machine ignores.
Frequently asked
Does ChatGPT recommend good consultants?
It recommends cited consultants, which is not the same thing. The model summarizes the firms named across comparison roundups and community threads. A firm can be excellent and invisible to AI because nobody has written about it, and a firm can be mediocre and highly recommended because it sits on the right listicles. Treat the recommendation as a reading list, not a shortlist.
Why does the same firm keep coming up in AI answers?
Because it is cited on the pages the model already trusts for that question. AI answers to vendor-selection queries lean on a small set of comparison roundups and threads. A firm named across several of them shows up again and again, regardless of whether it fits your stage or problem. Repetition in the answer reflects repetition in the sources, not consensus about quality.
How do I vet a firm that AI recommended?
Ask the AI where the recommendation came from, then read the original source. A named client, a piece of published work, or a specific result carries more weight than a spot on a paid or generic roundup. Then judge the firm on evidence you can verify: work you can inspect, references you can call, and a method the firm can explain. The AI cannot do that judgment for you.
Should I hire the firm ChatGPT names first?
Order is not ranking. In this category the field is fragmented, no single firm dominates, and the sequence shifts with the sources the model happened to pull. The firm named first is often just the one most written about, which correlates with marketing reach more than fit. Use the list to find candidates, then run your own diligence before you shortlist.