How We Tested This
We built 30 real queries about addiction treatment and ran each one twice on ChatGPT, a few days apart, logged out with no chat history or personalization. The queries varied by urgency framing (crisis-phrased, research-phrased, near-me), geography (Los Angeles, Atlanta, Nashville, Phoenix, Denver, Columbus, Bozeman), and substance/modality context. That gave us 60 logged runs. We screenshotted every result, logged the facilities named in order, and pulled Semrush Authority Score, referring domains, and backlink counts for every facility named.
WHAT THIS PIECE COVERS
This is a study of visibility: what ChatGPT actually recommends, how consistently, and based on what sources, across 60 real, logged runs. Every stat here comes straight from that log.
Full query list, tracking structure, and data sources are in the methodology appendix at the bottom, for anyone who wants to check our work. If you want the fundamentals on how AI search is already reshaping healthcare visibility, our GEO and AI Mode in healthcare guide covers that shift in depth.
Real Examples From the Data
These are real, unedited screenshots pulled directly from our logged runs, not mockups.
ChatGPT map results for an outpatient alcohol treatment question in Los Angeles, Run 1, showing one set of facility pins clustered around Santa Monica and Culver City
Same question, worded identically, asked days apart. The neighborhood, the pins, and the facilities on the map all changed. This is Finding 1 in practice, not a hypothetical.
ChatGPT response to a crisis-phrased query about a son addicted to opioids, giving overdose safety steps and hotline guidance instead of naming a facility
This is the typical crisis-mode pattern from Finding 2: safety guidance first, no facility name.
ChatGPT map results for a crisis-phrased Denver detox question, showing facility pins and ratings instead of a hotline redirect
This is the one real exception described below: a crisis-phrased query that still got a direct facility answer.
ChatGPT response to a calm question about VA addiction treatment options in Phoenix, naming specific facilities with Veterans Affairs source citations attached
This is what Finding 5 looks like up close: the citations attached to a real answer are government sources, not the facilities’ own marketing pages.
The Data Breakdown
Finding 1: The Top Pick Doesn’t Hold Still
Across the queries we ran twice, the top pick’s stability breaks down like this:
- The #1 recommendation matched between Run 1 and Run 2 only 43% of the time.
- On the identical question, worded the same way, a few days apart, the top answer changed 57% of the time.
Organic search results tend to hold steady for a query over a similar window. This doesn’t. It behaves more like a fresh guess each time.
Finding 2: Crisis Language Shuts the Door on Facility Names
- Queries phrased with crisis language triggered a safety or crisis-mode redirect (hotlines, general safety guidance, no specific facility) in 14 of 16 runs (88%).
- Calm research-style questions and near-me lookups triggered that same redirect in 0 of 44 runs.
Bar chart comparing how often ChatGPT triggered a safety-mode redirect: 88% for crisis-phrased queries versus 0% for calm-phrased queries
ONE REAL EXCEPTION
Only one crisis-phrased query broke this pattern, in both of its runs: an opioid-crisis question naming Denver directly. Instead of a hotline redirect, ChatGPT named a specific facility, Denver CARES, both times. It’s a single case, not a new rule, but it shows the crisis-detection behavior isn’t a hard filter on certain words. It’s reading the whole query, and a named city plus specific substance language was apparently enough to tip it toward answering directly.
Finding 3: Location Barely Factors In
- The model asked the user for their location before answering in only 4 of the 60 runs.
- The rest of the time it answered generically or off whatever city was already in the question.
Finding 4: Drug Rehab SEO Doesn’t Cleanly Predict the Winner
We pulled Semrush Authority Score for every unique facility recommended, split by #1 pick (F1) versus #2/#3 pick (F2/F3):
| Group | Facilities (n) | Mean Authority Score | Median | Share with AS ≥ 30 |
|---|---|---|---|---|
| #1 recommendation (F1) | 32 | 29.6 | 29.5 | ~50% |
| #2/#3 recommendation (F2/F3) | 59 | 30.4 | 29 | ~49% |
Bar chart comparing average domain Authority Score for the #1 recommended facility versus #2/#3 recommendations, with range bars showing both groups span 8 to over 60
The two groups are statistically close to indistinguishable. The facility landing in the #1 spot doesn’t carry meaningfully more domain authority than the ones landing second or third.
Across all 91 recommendation instances (32 in the #1 spot, 59 in #2/#3), the average Authority Score is right around 30, and the range is wide: from 8 up to 72. Most recommended facilities do carry a real, if modest, backlink profile, but it’s far from a requirement. AriseVida Care (Authority Score 8), Thrivur Health (9), and True North Recovery Services (17) all landed in the recommendation set despite domain authority that low.
PRO TIP
One clarification before drawing conclusions: what we measured here is specifically backlink-based domain authority, which is only one part of SEO, not the whole picture. SEO also covers things like certifications and accreditations, review volume and quality, and how consistently a business is mentioned across other trusted sites, and none of that is captured in an Authority Score. Backlinks alone don’t cleanly separate the #1 pick from #2 or #3: authority scores as low as 8 and 9 still made the recommendation set, and most facilities getting named do have some real backlink profile, just not a strong or consistent one. So the fair conclusion isn’t “SEO doesn’t matter.” It’s that backlink authority specifically looks like a soft baseline rather than a scoreboard, which lines up with what outside research on AI search is also finding: independent analyses of ChatGPT’s recommendation behavior report that traditional signals like backlinks and domain authority carry little weight compared to third-party mentions, reviews, and credentialing (see Finding 5 below for what that looks like in this specific dataset).
Finding 5: ChatGPT Leans on Government Health Data, Not Facility Marketing
We also logged every source ChatGPT cited alongside its recommendations.
- Of the 46 runs that named at least one facility, 42 (91%) cited a government or public-health authority somewhere in the answer, sources like SAMHSA, a state behavioral health agency, county public health, FindTreatment.gov, VA.gov, CDC, or MedlinePlus.
- Only a minority of runs cited the facility’s own website directly or a commercial treatment directory, and 3 runs cited no source at all.
Bar chart comparing how often ChatGPT’s facility-naming answers cited a government or public-health source versus the facility’s own site or nothing at all
This tracks with broader research on how AI answer engines work outside of addiction treatment specifically. Independent GEO analyses have found that third-party mentions, reviews, and credentialing tend to carry more weight in AI recommendations than a business’s own site or backlink profile. Our facility-level data backs that up: the thing that shows up most in ChatGPT’s citations here isn’t a treatment center’s own marketing, it’s a government or public-health source describing that facility.
PRO TIP
This is a real, practical lever, and it’s one most treatment centers aren’t managing at all. If ChatGPT is pulling public-health sources into its answers 9 times out of 10, then how accurately and completely your facility is represented in those government and public-health datasets, the SAMHSA treatment locator, your state’s behavioral health directory, your county’s public health listings, plausibly matters as much to AI visibility as your own website’s SEO. That’s a checklist item most marketing plans don’t even include yet.
What This Means for Treatment Center Marketing
A few practical takeaways follow directly from this data, whether you run addiction treatment marketing or behavioral health marketing more broadly.
Being Unmentionable Is the Real Risk, Not Being Unranked
Since domain authority barely separates who lands first from who lands third, and the top answer itself isn’t stable, chasing the #1 spot in an AI answer means chasing something that moves on its own every few days. The more durable goal is enough consistent, well-cited presence across the web (your own site, directories, press mentions, local citations) to be in the consideration set at all.
Local SEO for Rehab Centers Still Matters, Just Not the Way You’d Expect
Since the model rarely confirms location with the user, whatever already dominates local search visibility in a market tends to surface for near-me-style questions. Keep local SEO fundamentals solid, not because the AI is doing sophisticated location matching, but because it largely isn’t. Our guide on local SEO for treatment centers walks through how to win the “rehab near me” search specifically.
Get Your Facility Right in Government and Public-Health Listings
Since 91% of runs that named a facility cited a government or public-health source, keeping your listing accurate and current in the SAMHSA treatment locator, your state’s behavioral health directory, and county public health resources is now a visibility channel in its own right, not just a compliance checkbox.
Generative Engine Optimization Is Now Part of the Job
The category name for making sure AI answer engines can find, understand, and surface your facility accurately is generative engine optimization (GEO), and this data suggests it behaves differently from traditional SEO: less positional stability, less clean separation between top spots, heavy sensitivity to how a person phrases their question.
This is exactly where Pinnacle Pursuit comes in. We build addiction treatment marketing and behavioral health marketing programs designed for how people actually search now, across Google and ChatGPT alike, so your facility has the consistent, well-cited presence this data shows actually matters.
About This Sample
This is the pilot round of an ongoing series, and the scope was deliberate.
- 30 queries, run twice each. That’s enough real data to surface a clear, repeatable pattern (the instability finding held up consistently across the set), and it’s the sample size we’re building on for the next, larger round.
- Five major treatment markets, plus two smaller ones. Los Angeles, Atlanta, Nashville, Phoenix, and Denver anchor the dense-market data, with Columbus and Bozeman as an early look at smaller markets. A future round will expand that mix to say more about rural-versus-urban differences.
- Every facility we compared had a verifiable public domain. A handful of near-me names didn’t, so we left them out of the authority-score comparison rather than guess.
- This piece is a visibility study. It shows what ChatGPT actually recommends and why, straight from 60 logged runs.
- This round focused on ChatGPT. It’s the natural place to start, and the same query matrix is built to extend to other AI answer engines in a follow-up.
Methodology Appendix
Tools Tested
ChatGPT, consumer app, logged out, no personalization or chat history.
Query Design
30 queries varying urgency framing (crisis / research / near-me), geography (Los Angeles, Atlanta, Nashville, Phoenix, Denver, Columbus, Bozeman), and substance/modality context.
Execution Protocol
Each query run twice, several days apart, screenshotted, with the facility names, order, and any safety/crisis-mode redirect logged for each run.
Backlink and Authority Data
Pulled from Semrush Domain Overview for every unique facility domain named across all 60 runs, recording Authority Score, referring domains, and total backlinks.
Source Citation Analysis
For every run that returned at least one facility name, we logged every source ChatGPT cited in its answer, then classified each run by whether it cited a government or public-health authority (SAMHSA, a state agency, county public health, a federal resource like VA.gov or FindTreatment.gov), the facility’s own site, a commercial directory, or no source at all.
Data Availability
The full query log, screenshots, and Semrush pulls are available in our public tracking sheet for anyone who wants to check the numbers themselves.
Frequently Asked Questions (FAQs)
No. Asking the identical question twice, a few days apart, produced a different #1 recommendation 57% of the time in our test. There’s no fixed “top spot” to win here, so the smarter goal is staying consistently in the rotation rather than chasing #1.
Rarely. It asked for location before answering in only 4 of 60 runs we logged. That means having your location and service area clearly represented everywhere matters more, since ChatGPT usually isn’t double-checking it fits.
The model shifted into a safety or crisis-mode response, hotlines and general guidance instead of a specific facility, in 88% of crisis-phrased queries, versus 0% for calmly phrased ones. That means facility visibility mostly reaches the calmer, research-phase searchers, not people in acute crisis.
Around 30 out of 100, across the 91 recommendation instances we logged, with a wide range from 8 to 72. That spread shows domain authority isn’t gatekeeping who gets recommended, so it’s one piece of visibility, not the whole strategy.
Mostly government and public-health sources. 91% of runs that named a facility also cited something like SAMHSA, a state behavioral health agency, or FindTreatment.gov, rather than the facility’s own website. That makes those public listings a real visibility channel, not just a compliance item.
Four things matter most based on this data: keep your facility accurately listed in government and public-health directories, solidify your local SEO fundamentals since ChatGPT rarely confirms location, build a consistent well-cited presence across the web rather than chasing the #1 spot, and treat AI visibility as ongoing work, not a one-time fix.
