Issue 004 · November 2026For founders, marketers & owner-operators · monthlyAll issuesSign in

Fewer clients hear about agencies from a friend. Get your referrals in writing

Only 35% of clients now hear about new firms from friends or co-workers, down from 60% in 2022. AI assistants read what your clients write about you instead. Here's how to ask for it.

An old brass speaking tube on a dark wall above a neat stack of blank cream index cards, lit by a warm lamp
Image generated by AI, screened in two inks

Fewer clients now find a new agency through someone they know. RSW/US, which sells outsourced new-business programmes to small and mid-sized US agencies, says that by 2025 just 35% of clients learned about new firms from friends or co-workers. In 2022 it was 60%. Networking fell from 73% to 58%, and past relationships from 67% to 48%.

RSW/US has a reason to say referrals are drying up, and its 2026 outlook announcement doesn't give a sample size. But the direction matches what buyers say they now do instead.

“The assistant can't ring your old client. It reads what they wrote.”

The good word about you hasn't stopped. It's moved. When a marketing director asks ChatGPT for "a branding agency that has worked with B2B fintech", the assistant can't ring your old client. It reads what they wrote.

So the referral ask needs an update. "Mention us if anyone asks" still works, a bit. A client who writes it down, in public, with the details a buyer would type into a chatbot, keeps referring you while you sleep.

What buyers and assistants now do

Agencies are what B2B buyers research with AI most. Semrush surveyed 643 US B2B professionals in March and April 2026. Asked what they research using AI tools, 51% said "agencies and service providers", top of the list. Semrush sells tracking of how brands appear in AI answers, so it has a stake in the result. Three findings matter to a ten-person firm:

  • Only 7% notice a vendor in an AI answer because they recognise the name. 53% notice one that closely matches their specific use case.
  • After an AI mentions a vendor, 71% visit its website and 38% check reviews on G2 or a similar platform.
  • Just 14% ask colleagues. The friend who used to make the intro is now a minority step.

The assistants lean on other people's words. Muck Rack, which sells AI and PR software to comms teams, looked at more than 25 million links cited by ChatGPT, Claude and Gemini for its May 2026 report. 84% of citations came from earned media, a broad bucket covering journalism, academic and government sources, encyclopaedias and third-party company content. Paid and advertorial content was 0.3%.

The directories have noticed. On 23 April, Clutch put an app inside ChatGPT that surfaces "verified profiles, review highlights, pricing signals, and real client feedback" when a buyer asks for a provider. On 5 May it gave listed firms a free dashboard showing how often they appear in ChatGPT, Claude, Perplexity and Gemini answers. Its advice for doing better includes "collecting reviews from trusted sources".

The play: get your referrals in writing

  1. Pick ten. List the last ten clients who'd happily recommend you. If you can't find ten, start with the ones who renewed.
  2. Pick two places. Your Clutch profile, plus the one place your buyers already look: Google for a local firm, G2 if you sell to software teams, or a LinkedIn recommendation on the founder's profile. Two well-stocked profiles beat six empty ones.
  3. Ask for a specific review, not a nice one. "Great to work with" matches no query. "They rebuilt onboarding emails for our 40-person fintech and trial-to-paid went up" matches the use case a buyer types, which is what 53% of them notice.
  4. Put the same facts on your own site. In the Semrush survey, 71% went to a vendor's website after an AI named it. A case study whose first line gives sector, size, problem and result confirms what the assistant said.
  5. Make it routine. Ask at a natural high point: the end of a project, a renewal, the first good results report. Once per client per year is plenty.

Here's the ask we'd send. Subject line: "Could you write three lines about us?"

"Hi Sam, we're asking a few clients for a short review on Clutch (link below). More buyers now find agencies by asking AI tools, and those tools read reviews. If you have five minutes, the most useful ones answer three questions. Your words, not ours, and no pressure if now's not a good time."

  • What kind of company are you, and what was the problem?
  • What did we do?
  • What changed? A number helps, if you're happy to share one.

The cost and the risk

You're spending goodwill. Ask once, make it easy, and skip the client who's mid-wobble on a deliverable.

Don't write it for them. In the UK, fake reviews and reviews that hide an incentive have been banned since 6 April 2025 under the Digital Markets, Competition and Consumers Act. The government's guidance names "writing or commissioning others to write fake reviews" as banned, and the CMA can fine up to 10% of annual global turnover. The rules are framed around consumer reviews; treat them as the floor anyway. Suggest the questions, never the answers. If you offer a thank-you, make sure the review says so.

Directories sell upgrades. Clutch says profiles with its Verified status average "8X more AI citations than free profiles". That's Clutch's number, and the announcement doesn't say how it was measured. Fill in the free profile and collect reviews first. Watch whether the dashboard moves before you pay for anything.

And nobody has shown that ten reviews buys a place on a ChatGPT shortlist. The evidence says assistants read third parties and buyers check reviews. It doesn't say how many it takes. If referrals still bring in most of your work, this is a quarter's slow project, not a fire.

This week

Open a doc, list your ten clients, and send the first three asks by Friday.

Also in this issue