Your Next Referral Will Come From a Machine

Your next referral will not come from a dinner party. It will come from a language model answering "who should I see for mold toxicity near me." I have spent twenty years writing the software that decides what gets shown to whom, and I am telling you the judge changed jobs. The judge used to be a neighbor who liked your bedside manner. Now it is a model that reads your website, your reviews, and your schema, and it does not care how warm your handshake is.

This should terrify anyone still treating their online presence like a digital business card. It should also excite you, because the new judge is more honest than the old one. It cannot be charmed at a dinner table. It can only be shown evidence.

The Position That Used to Require a Dinner Table

The most valuable position in any market has always been the trusted advisor, the person other people send their friends to without being paid for it. For most of medical history that position was earned slowly. You built it one patient at a time, one referring physician at a time, one conversation at a time. It compounded, but it compounded on human memory, and human memory is small and local.

I have watched this exact dynamic play out with Dr. Piper Gibson. She is excellent at what she does. For years that excellence was invisible past her own patient list, because the only distribution mechanism was word of mouth and a website that nobody could find. We rebuilt her visibility from the ground up: structured content, clear entity signals, a site organized around the actual questions her patients ask. She went from essentially unfindable to ranking for the terms her ideal patients search. Nothing about her clinical skill changed. What changed is that the systems deciding who gets found could finally read what she does.

That is the whole story of this essay in miniature. The clinical work was never the constraint. The constraint was legibility to the system doing the referring.

The New Referral Engine Runs On Text, Not Memory

When someone types a health question into ChatGPT or Perplexity today, the model is doing something structurally similar to what a well-connected friend used to do: synthesizing everything it knows and naming names. But the model's memory is not a dinner party. It is a corpus. It reads your site's structured data, the pattern of language across your reviews, how deep your content goes on the specific conditions you treat, and whether other sources on the internet describe you consistently.

None of that is subjective in the way a friend's impression is subjective. It is closer to a credit score. Inputs go in, a ranking comes out, and the ranking is legible if you know what to check.

Here is the part most practices get backwards: they assume if they show up well on Google, they show up well to the models. Those are different systems reading different signals. Google still leans on links and behavior data. The models lean harder on structured, self-consistent, deeply specific content. A practice can rank fine on Google and be a ghost to Perplexity. I see this constantly in functional medicine, where practices write generically about "hormone balance" and "gut health" and never go deep enough on the actual entities, the actual protocols, the actual conditions, for a model to confidently attach their name to a specific question.

The Algorithmic Referral Audit

Run this today. It takes fifteen minutes and costs nothing.

Step 1: Ask the machines about yourself. Open ChatGPT, Perplexity, and Google's AI Overview separately. For each one, type variations of:

1. "Who is the best [your specialty] doctor near [your city]?"
2. "I have [your top three conditions you treat] symptoms, who should I see in [your city]?"
3. "Compare functional medicine practitioners in [your city] for [specific condition]."
4. "What is [your practice name] known for?"

Step 2: Grade what comes back. Look for four things, in this order:

• Do you appear at all? If your name or practice never surfaces across ten to fifteen prompt variations, you are invisible to this channel entirely. That is the most urgent finding.
• Is what it says accurate? Models sometimes hallucinate specialties you don't offer or drop ones you do. Wrong information is worse than no information, because a referred patient shows up expecting the wrong thing.
• Is it specific or generic? "A well-reviewed functional medicine practice" is a non-answer. "Treats mold-related chronic illness using environmental testing and a phased detox protocol" is a real answer. Specificity is what makes a model confident enough to recommend you by name.
• Is it consistent across models? If ChatGPT describes you one way and Perplexity differently, your underlying signals are inconsistent, and inconsistency reads as unreliable data to systems trained to be cautious with health claims.

Step 3: Trace the answer back to its source. Ask the follow-up: "Where did you get that information?" Perplexity in particular will often cite sources. Follow those links. You are looking for whether your own site, your directory listings, and your review platforms are telling the same story in the same words.

The Four Signals That Actually Decide the Answer

Once you know where you stand, here is what to fix, in priority order.

1. Entity consistency. Your practice name, your specialty language, your location, and your provider name need to match, word for word, across your website, Google Business Profile, directory listings, and any bios you've published elsewhere. Models cross-reference these. Mismatches don't just confuse patients, they weaken the model's confidence that it has found one coherent entity worth recommending.
2. Topical depth, not topical breadth. Stop writing one page that mentions ten conditions shallowly. Write one page per condition that goes deep: mechanism, your specific approach, what a first visit looks like, what the testing involves. This is Visibility Leverage in practice: you are not trying to rank for everything, you are trying to be the unambiguous best answer for a narrow, specific question.
3. Structured data that describes what you actually do. Schema markup on your site, your services listed as distinct entities rather than a paragraph of prose, FAQ sections that answer the literal questions patients type into search bars. This is the plumbing most practices skip because it's invisible to human visitors. It is not invisible to the systems reading your site to decide who to recommend.
4. Review language in the patient's own words. Models weight review corpora heavily because reviews are the closest thing to unbiased evidence they have. If your reviews all say "great doctor, very caring," that is warm and says nothing. Reviews that mention the specific condition, the specific protocol, and the specific outcome are the raw material that builds Authority Leverage: third-party language a model can quote back as evidence, not just an assertion you made about yourself.

What To Fix First

If you only do one thing this week, run the audit above and read what the machines already say about you. You cannot fix a signal you have not measured. Most practices are optimizing a channel, human word of mouth, that is shrinking in relative share, while ignoring a channel that is actively deciding referrals right now, invisibly, every day.

The honest version of this shift is that the work did not get harder. It got more legible. The practice that documents its specific expertise clearly, consistently, and in depth wins this new channel the same way the practice with the best local reputation used to win the old one. The mechanism changed. The requirement to actually be excellent and to actually say so clearly did not.

Run the audit. Fix entity consistency first, it's usually a few hours of cleanup. Then go deep on your top three conditions. Everything else follows from those two moves.

If you want a second set of eyes on what the audit turns up, that's what I do. Work with me.