thought experiment...

Channel / marketing shift · September 2026

You’re no longer competing to be found. You’re competing to be recommended.

A search page showed ten options. An AI answer compresses that to two or three names, with a reason attached. Fourth place is close to invisible.

Where things stand

Fourth place is a very polite exit.

01 / where things stand

The short version.

Four things worth keeping, and a path that doesn’t show up in the analytics. Don’t chase the traffic number.

  1. 01 / Adoption

    45%

    from 6%, in a year

    BrightLocal, March 2026. Local AI recommendations went from 6% to 45%. AI is now the third discovery channel, ahead of Yelp.

  2. 02 / Referrals

    1.4%

    lowest AI referral share

    Local services sit at the bottom. Maps and proximity still live outside the assistants, so a Google Business Profile still matters.

  3. 03 / Conversion

    42%

    better. A year earlier, 38% worse.

    Adobe, March 2026. AI-referred visitors converted 42% better. A year earlier, they converted 38% worse. The recommendation is doing the selling before the click.

  4. 04 / The invisible path

    Askbrandbuy

    Credit lands on the search.

    Ask an assistant. Search the brand on Google. Buy. The dashboard thanks branded search. The choice was made one step earlier.

Who’s a good therapist in Vancouver for an adult with ADHD who works in tech?

A search page. Ten options. No reason attached.

Ten search results, no reason attached

  1. Counsellor in Vancouver
  2. Therapist near me
  3. Anxiety counselling, downtown
  4. Couples therapy Vancouver
  5. ADHD therapist
  6. Counselling for professionals
  7. Best therapists in Vancouver
  8. Walk-in counselling
  9. Online therapy, B.C.
  10. Counsellor accepting new clients

Two or three names, with a reason

  1. In the answer

    A name

    Adult ADHD, people who work in tech, Vancouver. The reason matches the question, and another source agrees.

  2. Also in the answer

    A second name

    Same niche. A different detail you could check: training, a directory, a piece they actually wrote.

  3. Still named

    A third name

    Adjacent. A weaker fit. Still in the answer.

Everyone else on the list of ten. Still published. Not in the decision.

Ten names is a page. It is not yet a recommendation.

A sketch of the shape of an answer. Not a ranking. No clinic was placed here.

02 / what we’d actually do

Five principles. None of them a plugin.

Be the best-described, most-corroborated option. Then make that description easy to repeat.

A recommendation is farm-to-table. It has to come from somewhere real.

  1. Be a consistent entity, not just a website.

    The same name, the same specialties, the same address, the same one-liner. On the site, the Google Business Profile, the directories, the podcasts, the articles. Third-party mentions carry weight a homepage can’t. As of March 2026, Reddit is the first or second most-cited source on major AI engines.

  2. Write for the questions people ask before they need you.

    An assistant fans one question out into smaller ones. Answer in the first paragraph. If the useful sentence is buried under the origin story, it may never travel.

  3. Publish what nobody else can.

    Original research. A framework with a name. Commentary only you could have written. A generic list of tips is close to worthless. The model already has a thousand of those, and they sound like each other.

  4. Make yourself machine-legible and machine-bookable.

    Schema. Pricing in plain language. A way to book that doesn’t require a scavenger hunt. Agents that book are coming. If a person can see what you do and what it costs, a machine has a chance too.

  5. Measure it by hand.

    Once a month, ask ChatGPT, Gemini, Claude, and Perplexity the questions your ideal client would actually ask. In a study of 2.05 million local questions, four engines agreed on the top business only 4% of the time. Add “an AI assistant” to the intake line “how did you hear about us.” If you don’t ask, the recommendation never shows up in the report.

03 / a constraint, not a loophole

The therapy wrinkle.

Corroborated, not solicited.

Reviews are a major trust signal. Asking for them may not be allowed.

Soliciting testimonials may be restricted, and for therapists it often is. Before anyone asks a client for a line, check the BCACC standard on marketing, advertising, and fees (opens in a new tab).

The substitutes are slower, and they’re the ones that hold up: professional citations, guest articles, directories, and writing that shows the expertise instead of announcing it.

04 / under the brief

Questions most people aren’t asking.

The useful ones show up before a proposal, not after a dashboard.

If you can’t say the sentence, the model will.

  1. What does an assistant say about you when you’re not in the room — and is any of it wrong?

    A mismatched specialty. An old address. A service you retired. Wrong details travel. If you haven’t looked, you don’t know which version of you is being named.

  2. What is the one sentence — and is it distinct?

    “Counsellor in Vancouver” loses. A specific niche and a specific background wins. If you can’t say the sentence, the assistant will invent a duller one.

  3. If the information is free, what are you charging for?

    Judgment. The relationship. Accountability. An assistant can summarize the field. It cannot sit with someone and be responsible for what happens next.

  4. Could a practice sit on this gap?

    Helping clinicians become the recommended answer. It overlaps an advertising background with Magnusson-style builds: a person made legible, bookable, and corroborated, without being turned into an ad.

  5. Are you building on a trick that will get patched?

    Self-promotional listicles — about one in ten citations from late 2025 into early 2026 — get corrected. Build on being the best-described option, and the one other sources agree about.

05 / Byron’s note

Shift where the return comes from.

Days on the idea. Weeks on the work.

Television, sponsorships, and generic SEO were reasonable ways to buy attention. They don’t signal proficiency to an assistant the way information does, or lived experience, or a niche you can actually describe, or proof that only you could have published.

The old levers still spend money. They spend it somewhere the recommendation can’t see. Shift where the return is influenced from.

The collective’s bias applies here, lightly. Build it. Break the version that doesn’t get named. Learn what the assistants actually say, then build the next one. Days on the idea. Weeks on the execution. Not months on a proposal for a channel you haven’t checked.

Good brains. Less baggage. The baggage, here, is a plan nobody can recommend.

06 / where the figures live

Sources.

The figures above come from these reports. A number that isn’t in the piece isn’t one we used.

  1. Peec AI GEO statistics 2026 peec.ai(opens in a new tab)
  2. LovedByAI State of AI Traffic for Small Businesses 2026 lovedby.ai(opens in a new tab)
  3. First Page Sage AI Impact on Website Traffic 2026 firstpagesage.com(opens in a new tab)
  4. Digital Applied AI Shoppers Convert 42% Better digitalapplied.com(opens in a new tab)
  5. LLMrefs GEO 2026 Guide llmrefs.com(opens in a new tab)
  6. Search Engine Land Mastering GEO in 2026 searchengineland.com(opens in a new tab)

A thought experiment

Make the true thing
easy to recommend.

Then check whether anyone did. The people you talk to are the people doing the work.

Talk to the collective

Sample channel piece · September 2026