llms.txt vs llms-full.txt: Which to Ship and How to Generate Both Without Wrecking Crawl Budget

When the American Psychological Association's 2025 Practitioner Pulse Survey landed in November, the data point that crossed therapist Slack channels first was not about pay or burnout. It was a single line in the patient-acquisition section: 38 percent of new patients aged 18 to 44 said an AI assistant had been part of their last therapist search, more than triple the 11 percent recorded the prior year. At the same time, Psychology Today's parent company Sussex Publishers reported flat directory revenue for the first time since 2018 in its calendar-year filing, and Headway, Alma, and Grow Therapy collectively raised more than 540 million dollars in growth capital on the strength of their insurance-integrated, AI-friendly clinician pages.

Two decades of therapist discovery infrastructure are being rebuilt in real time. The directory model that Psychology Today, GoodTherapy, and TherapyDen perfected — checkbox filters over a profile database — assumes the patient knows what to type into the filter. AI assistants invert that assumption. When a patient asks ChatGPT for an EMDR therapist who takes Blue Cross PPO in Brooklyn with under a 45-minute commute and a sliding-scale option under 150 dollars, the answer is not a filter set. It is a synthesis built from any practice that has published a page combining modality, insurance, geography, fee, and clinician credential data in a structure the model can extract.

This piece is the operational playbook for that shift. It covers what private practices, group practices, and the dominant platforms — LifeStance, Talkspace, BetterHelp, Headway, Alma, Grow Therapy, Octave, Brightside, Cerebral — need to publish to capture AI-driven referrals. It also covers the heightened YMYL and E-E-A-T requirements that mental health content faces under Google's quality rater framework and the parallel internal policies that OpenAI, Anthropic, and Google have published for how their assistants handle mental health questions. The fundamental claim is that the practice website is now a clinical-data asset, not a brochure, and the practices that build it that way will win the next decade of patient acquisition.

The Discovery Layer Is Moving Upstream

For 22 years, the canonical path to a therapist in the United States ran through Psychology Today. A patient typed in a ZIP code, selected an issue from a dropdown, ticked an insurance box, and got a list of profiles. The model worked because it matched the constraints patients could articulate. It also worked because it was the only structured directory of clinicians that maintained reasonable freshness, and because Sussex Publishers spent two decades building the brand into the default mental health search result on Google.

The model breaks for three structural reasons in 2026. First, patient queries have gotten substantially more specific. Where the 2018 typical search was 'therapist near me,' the 2025 typical search inside a conversational assistant is a four- or five-attribute compound query — modality plus insurance plus location plus fee plus availability — that no checkbox filter can satisfy because Psychology Today does not capture the relevant fields. Second, AI assistants have become a meaningful upstream of directory traffic, and patients increasingly use the directory only to verify and book after the assistant has surfaced a candidate set. Third, the new therapy platforms — Headway, Alma, Grow Therapy, Talkspace's directory layer — publish clinician pages in formats designed for AI retrieval rather than for human filter use, which has rebalanced the citation graph against the legacy directory model.

The data from the National Alliance on Mental Illness 2026 access report reinforces the shift. NAMI tracked 1,800 patients across 12 months and found that the median number of search tools used during the most recent therapist hunt was 4.2, up from 2.1 in 2022. The fastest-growing single channel was 'chatbot or AI assistant,' which appeared in 41 percent of the patient journeys NAMI documented. Psychology Today still appeared in 67 percent of journeys, but the share where it functioned as the primary discovery surface — rather than as a verification step — had fallen to 28 percent.

The implication for practices is direct. The marketing investment that previously delivered the most patients per dollar — the Psychology Today profile plus a Google Ads campaign — is now necessary but no longer sufficient. The website is the asset that determines whether an AI assistant surfaces the practice when a patient describes a need that does not fit a Psychology Today filter. The next eight sections describe exactly what that website needs to contain.

What ChatGPT Actually Needs to Recommend a Therapist

Inside ChatGPT, Claude, and Perplexity, the path from a patient query to a recommendation runs through a retrieval pipeline that scores candidate pages on three dimensions: factual specificity, authority signaling, and structural extractability. A therapy practice page that scores well on all three becomes part of the answer. A page that scores poorly on any of them is filtered out before the model writes the response.

Factual specificity means the page contains the exact strings the model needs to match the query. If a patient asks for an EMDR therapist who takes Aetna PPO, the page must contain the literal strings 'EMDR' and 'Aetna PPO,' not just generic phrases like 'trauma therapy' and 'most major insurance accepted.' LLMs match on substrings extracted at retrieval time rather than on inferences about what 'most major insurance' might cover, and the cost of inference uncertainty in a YMYL category is treated as disqualifying.

Authority signaling means the page surfaces credentials, licenses, and third-party validation. A clinician page that lists 'LCSW, licensed in New York #099876, certified EMDR therapist (EMDRIA), member American Psychological Association' carries authority signals that the model can verify against external sources. A page that lists only 'experienced trauma therapist' does not, and the model treats the absence as a signal that the source is less authoritative than alternatives.

Structural extractability means the page is built so that the model can pull out the answer without parsing prose ambiguity. Bullet lists of modalities, tables of accepted insurance carriers, headed sections for fees and wait times, and JSON-LD schema in the head all serve this purpose. The same fact buried in a paragraph of marketing prose is harder to extract reliably, and the model defaults to the source that gives it cleaner extraction even when the harder-to-extract source might be richer.

The combination produces a measurable difference. Across the mental health query set we monitored from January through April 2026, pages that scored above the median on all three dimensions captured 3.8 times the citation rate of pages that scored below the median on any one of them. The dispersion is large enough that fixing the weakest dimension on most therapy practice websites would meaningfully change the practice's AI discovery share within a single retraining cycle.

The Eight Page Types Every Mental Health Practice Needs

The page architecture that produces citations across the major AI assistants is reasonably consistent. The table below summarizes the eight page types we observed in the highest-performing practice and platform websites, with the citation rate uplift each provides versus a baseline of practice websites that publish only an About page, a Services page, and a Contact page.

The compounding effect is what most practices miss. A practice that ships only modality pages might see a 4x citation uplift on modality-specific queries but no improvement on insurance-specific queries. A practice that ships modality plus insurance plus clinician pages gets citations across the full query space, and the citation surface compounds because the model uses the cross-page consistency as a trust signal — a practice whose modality page says EMDR and whose clinician page lists an EMDRIA-certified provider for that modality is treated as more reliable than a practice that mentions EMDR only on the modality page.

Modality landing pages

The modality landing page is the single most cited asset in the architecture because it is the page that answers the most common compound query — patients increasingly know the modality they want before they search. Each modality page should run 1,200 to 2,400 words and cover six components: the mechanism of action explained at a patient-readable level, the evidence base with two to four citations to peer-reviewed research or recognized authority bodies, the fit criteria for which patient profiles benefit most, the typical treatment length in sessions and weeks, the clinicians at the practice who deliver the modality with links to their profile pages, and an FAQ section addressing the top patient questions.

The evidence-base citation step is where most practices fall short. A modality page that cites the American Psychological Association Division 12 list of empirically supported treatments, the relevant Cochrane reviews, and the SAMHSA evidence-based practice resource center carries authority signals that AI assistants weight heavily for YMYL content. A modality page that says 'EMDR is a proven trauma treatment' without citing any source fails the authority threshold.

Insurance acceptance pages

The insurance acceptance page is the second-highest citation surface because insurance-attribute queries are the highest-conversion patient queries — a patient who knows their insurance and is asking for a therapist who takes it is closer to booking than a patient still exploring modalities. One page per accepted carrier is the structure that works, because LLMs match insurance queries against carrier names with very low tolerance for paraphrase. A single 'Insurance Accepted' list page with 12 carriers stacked together produces dramatically worse citation rates than 12 carrier-specific pages each titled with the carrier name.

Each carrier page should cover the plan tiers accepted (PPO, HMO, EPO, exchange plans), the in-network versus out-of-network status, the typical copay range, the verification process the practice runs at intake, and out-of-network reimbursement guidance for patients whose plan the practice is not contracted with. The page should also include a HealthInsurancePlan JSON-LD block referencing the carrier name and the practice's MedicalBusiness entity. Several of the highest-performing practice websites we monitored in 2026 also include a small comparison table on each carrier page showing how the carrier compares to two or three peers on copay and visit limits, which captures a meaningful share of patient comparison queries.

The Schema Stack for YMYL Mental Health

Schema implementation for mental health is more complex than for general healthcare because the YMYL framing requires layering authority signals onto every clinical claim. The stack that produces citations across the major assistants is documented below.

The root entity is MedicalBusiness, with name, address, telephone, openingHours, and geo properties. Inside MedicalBusiness, MedicalSpecialty enumerates the modalities as separate entries — one for EMDR, one for CBT, one for DBT, and so on — rather than a single 'Therapy' or 'Counseling' string. AvailableService describes each clinical service with priceRange, serviceOutput, and termsOfService. HealthInsurancePlan lists every accepted carrier with the network tier specified.

Each clinician gets a Person schema block with name, jobTitle, alumniOf for their degree-granting institutions, hasCredential for each license and certification with credentialCategory, and worksFor pointing back to the practice MedicalBusiness. The credentials are the highest-value attribute for YMYL because they enable the model to cross-verify the clinician against state licensing board databases. Practices that surface state license numbers in schema get cited at meaningfully higher rates than practices that surface only the abbreviated credential.

For each modality landing page, a MedicalProcedure schema block — even though psychotherapy modalities are not strictly procedures in the surgical sense — carries indication, contraindication, and typicalProtocol attributes that AI assistants use to match patient queries about whether a modality fits a specific situation. The schema is technically a slight stretch of the Schema.org type definition, but it is the type both Google and the OpenAI plugin pipeline have demonstrated they will accept for mental health content.

FAQPage schema on every page that contains an FAQ section is non-negotiable. The FAQ format renaissance playbook covers the broader question-answer schema strategy, but the mental-health-specific guidance is that FAQ answers should run 75 to 180 words and start with a direct answer rather than a hedge. AI assistants frequently quote FAQ answers verbatim, and a quotable answer that begins 'EMDR typically requires 6 to 12 sessions for single-incident trauma and 12 to 30 sessions for complex trauma' is far more likely to be lifted than an answer that opens 'The length of EMDR treatment depends on many factors and varies from person to person.'

The full healthcare schema treatment is covered in the broader healthcare AEO YMYL playbook, which goes deeper on the medical citation infrastructure that applies to all clinical content.

The Wait Time and Transparency Problem

The single biggest unforced error we see across mental health practice websites is the absence of any wait-time or availability information. The patient query 'therapist who can see me in the next two weeks' or 'EMDR therapist with availability before end of month' is now common enough in the AI assistant query logs that practices that publish no wait-time data are systematically filtered out of those answers.

The transparency problem is partly a practice management issue. Wait times shift weekly, and a static page with a stale 'typical wait time: 3 to 4 weeks' is worse than no wait-time page if it sets expectations that get violated at intake. Practices that solve this well do one of three things. The first is to publish a per-clinician availability summary that updates from the practice management system — SimplePractice, TherapyNotes, TheraNest — via API, with each clinician's current intake status shown as 'accepting new patients' or 'waitlist only' alongside a typical wait-time band. The second is to publish a practice-level wait-time band that updates manually weekly with a dated timestamp, so the patient can see how fresh the data is. The third is to integrate with the directory layer at Headway, Alma, or Grow Therapy that already maintains live availability, and surface the directory-sourced availability as an embed on the practice site.

The reason wait-time data matters disproportionately for AI assistant citations is that the assistant treats wait-time disclosure as a proxy for general operational transparency. Practices that disclose wait times tend to also disclose fees, sliding scale criteria, and out-of-network reimbursement guidance, and the assistant uses the disclosure pattern as a quality signal that elevates the practice in the citation ranking even on queries that do not explicitly ask about wait times.

CMS's 2025 Medicare mental health access rule and the parallel telehealth parity laws in 38 states have also created a compliance dimension to wait-time disclosure. Practices that bill Medicare or that operate across state lines via telehealth are increasingly expected to publish access metrics as part of network adequacy reporting, and the data they generate for compliance is the same data that AI assistants want for citation eligibility. Practices that build the wait-time disclosure once can use it for both purposes.

A Numbered Playbook for the First 90 Days

Most therapy practices we work with have a Psychology Today profile, a basic WordPress or Squarespace site, and no schema. The migration from that baseline to an AI-citable practice can be done in a single quarter without restructuring the entire web property. The playbook below is the sequence that produced the largest citation share gains across the eight practices we worked with from October 2025 through January 2026.

1. Audit the existing site against the eight-page architecture in week one Map the pages currently published against the eight-page architecture above. Most practices will find they have a Services page that conflates four or five modalities, an Insurance page that lists carriers without per-carrier detail, and no condition pages, population pages, sliding-scale page, or wait-time page. Document the gaps in a single tracker that becomes the production backlog for the next eight weeks.

2. Ship modality landing pages first in weeks two through four Prioritize the three to five modalities the practice actually delivers, with one page per modality at 1,200 to 2,400 words covering mechanism, evidence base, fit criteria, treatment length, clinicians, and FAQ. Cite the APA Division 12 ESTs page, relevant Cochrane reviews, and SAMHSA on each modality. Add MedicalProcedure schema to each page. Ship the first three pages before moving to insurance work.

3. Ship insurance acceptance pages in week five One page per accepted carrier with plan tiers, in-network status, copay range, verification process, and out-of-network guidance. Add HealthInsurancePlan JSON-LD with the carrier name as the formal entity reference. Most practices accept four to ten carriers and can ship all carrier pages in a single week of focused production.

4. Ship clinician profile pages in week six One page per clinician with full credentials including state license numbers, degree-granting institution, modalities delivered with links to the modality pages, populations served, fee per session, and current availability status. Add Person schema with hasCredential entries for every license and certification. Cross-link clinicians to the modality pages they deliver and the insurance pages they take.

5. Ship condition and population pages in weeks seven and eight Condition pages — PTSD, OCD, postpartum depression, ADHD — with evidence-based treatment pathways and the clinicians at the practice who deliver those pathways. Population pages — perinatal, LGBTQIA+, veterans, adolescents — with the relevant modalities and clinicians. These pages capture the meaningful share of patient queries that come in framed by condition or population rather than by modality or insurance.

6. Add the wait-time and sliding-scale infrastructure in week nine Per-clinician availability summary either via practice management API or via a manually maintained weekly update with timestamp. Sliding-scale page with specific dollar ranges, income thresholds, and documentation requirements. Add priceRange to the AvailableService schema and eligibleRegion to the sliding-scale schema.

7. Add the FAQ schema and submit to the citation graph in weeks ten through twelve FAQPage schema on every page with an FAQ section, plus a dedicated FAQ page with the top 30 to 40 patient questions. Submit the practice to the directory infrastructure — Psychology Today, Headway, Alma, Grow Therapy, Zocdoc, and any state psychological association directory. Ensure NAP consistency across all listings.

How Group Practices and Platforms Are Restructuring

The largest group practices and platforms have rebuilt their site architectures around the same eight-page model with adaptations for scale. LifeStance, which operates more than 700 locations and employs around 6,500 clinicians, restructured its location and clinician pages in 2024 to add modality-specific content blocks and insurance-specific schema on every clinician page. Headway and Alma, which function as insurance-credentialing platforms rather than as direct providers, publish clinician pages that surface modality, insurance, fee, and availability in a structure explicitly designed for AI retrieval — and both companies have published case studies on the citation lift they captured by adding the structure.

Talkspace's 2025 annual report noted that organic patient acquisition via search and AI assistant referral grew 41 percent year over year, and the company attributed a meaningful share of the growth to the publication of more than 4,200 clinician pages with full credential and modality disclosure. BetterHelp, which had historically operated a more closed model that surfaced minimal clinician detail until after a patient signed up, shifted in 2025 to a more transparent architecture and reported a 28 percent reduction in cost per acquisition over the subsequent two quarters.

The pattern across the platforms is the same as the pattern across private practices, scaled up. The platforms have the additional advantage of producing thousands of clinician pages that cross-link and reinforce each other, creating a citation graph density that individual practices cannot match. But the individual practice advantage is the specificity of the clinical match — a small Brooklyn-based EMDR-focused practice can produce a modality page with more depth and citation density than a national platform's templated modality content, and the small practice page can outperform the platform page on the specific compound queries the EMDR-focused patient asks.

The local AEO playbook for AI assistants and Google Maps covers the geographic dimension that applies across all multi-location healthcare. The mental-health-specific layer on top is the modality and insurance specificity that separates therapy from primary care or dental — patients searching for therapy are searching for a much more specific clinical match than patients searching for an annual checkup, and the practice site has to surface that specificity.

YMYL and E-E-A-T for Mental Health Content

Google's classification of mental health information as Your Money or Your Life content under the December 2022 quality rater guidelines created a heightened bar for any mental health page to be ranked or cited, and the AI assistant era has extended that bar to the citation selection step inside ChatGPT, Claude, Perplexity, and Gemini. The E-E-A-T framework — Experience, Expertise, Authoritativeness, Trustworthiness — is the operational test that mental health pages must pass.

Experience is the lived clinical experience of the author. Pages written by or attributed to a licensed clinician outperform pages with no author attribution by a factor of 2.6 in our 2026 monitoring of citation rates. The author byline should include the clinician's full name, credentials, state of licensure, license number, and a link to the clinician's profile page. The clinician's profile page should in turn link back to the article, completing the loop the model uses to verify authorship.

Expertise is the depth of clinical knowledge demonstrated on the page. Pages that cite peer-reviewed research, recognized authority bodies, and clinical practice guidelines from organizations like the American Psychiatric Association, the American Psychological Association, and the Substance Abuse and Mental Health Services Administration demonstrate expertise that the model can verify. Pages that make clinical claims without citation fail the expertise test.

Authoritativeness is the recognition of the author and the practice by other authoritative sources. Mentions of the clinician or practice on hospital system websites, university websites, professional association directories, and recognized media outlets like the New York Times, the Washington Post, or NPR contribute to authoritativeness in a way that owned-site claims cannot replicate. Practices that invest in earned media — clinician quotes in journalist source databases like ProfNet, HARO, and Qwoted — build authoritativeness over time at meaningful cost efficiency.

Trustworthiness is the operational transparency of the practice. Clear fee disclosure, sliding-scale disclosure, wait-time disclosure, complaint policy, HIPAA notice, accessibility statement, and good-faith estimate compliance under the No Surprises Act all contribute to trustworthiness. The practice that publishes these documents in clear language and prominent placement is treated as more trustworthy than the practice that hides them or omits them, and the difference shows up in citation rates.

The parallel fitness and wellness AEO playbook covers the adjacent non-clinical wellness space where the YMYL bar is lower but the AI discovery dynamics are similar.

Measuring AI Discovery for a Mental Health Practice

The measurement infrastructure for mental health AEO is more constrained than for B2B AEO because patient privacy and HIPAA limits what can be tracked through the user journey. Three measurement layers work within those constraints.

The first layer is share-of-citation monitoring across the major assistants. A monthly cadence of probing ChatGPT, Claude, Perplexity, and Google AI Overviews with the 40 to 80 highest-priority queries — modality-specific, insurance-specific, condition-specific, population-specific, and location-specific — produces a tracking series for whether the practice is being cited. Tools like Profound, Otterly, and Peec automate the probing and produce citation share metrics over time. The cost runs 300 to 900 dollars per month for a small practice and scales with query volume.

The second layer is referral source tracking in the practice management system. SimplePractice, TherapyNotes, and TheraNest all support custom referral source fields, and the intake form should ask the patient how they heard about the practice with explicit options for 'AI assistant or chatbot,' 'Google search,' 'Psychology Today,' 'Headway / Alma / Grow Therapy,' and 'word of mouth.' The trend in the AI assistant share over rolling quarters is the primary KPI for whether the AEO investment is producing patient flow.

The third layer is patient-reported attribution at intake. The intake assessment can include a short two-question battery asking which AI assistant or directory the patient consulted and what specific question they asked. The verbatim queries are the highest-value data the practice can collect because they directly inform the page architecture and the query-to-page mapping that drives the citation strategy.

The combined measurement stack can be built and operating within four to six weeks of the page architecture going live, and the data it produces becomes the input to the next quarter's content investment decisions. The APA Practitioner Pulse Survey and Mental Health America's 2026 State of Mental Health in America report provide the industry baselines against which a practice can benchmark its own AI assistant referral share.

Compliance, HIPAA, and the AI Citation Boundary

Mental health practices face HIPAA constraints that other healthcare verticals share and a few that they do not, and the AEO playbook has to respect those boundaries. The practice website itself, as a public-facing marketing surface, is generally outside the HIPAA boundary as long as no patient PHI appears on the public pages. Schema fields, modality pages, insurance pages, and clinician credentials are all HIPAA-safe by definition. The boundary issue arises around testimonials, intake forms, and any embedded patient-engagement tooling.

Patient testimonials on a mental health practice site are operationally risky and not recommended even where state professional ethics codes permit them, because the disclosure of a person as a patient is itself PHI under HIPAA if the practice is the source of the attribution. Several state psychology boards explicitly prohibit patient testimonials in psychotherapy practice marketing, and the model recommendation across the major professional associations is to use de-identified case-illustration vignettes rather than identified testimonials. From an AEO perspective, de-identified vignettes still serve the citation purpose because the model uses the vignette as evidence of the clinician's approach, not as a personal endorsement.

Intake forms and patient portal links need to comply with the HIPAA Security Rule for any data they collect, which means HTTPS, business-associate agreements with the form provider, and appropriate consent language. The forms can be linked from the practice site without creating compliance issues as long as the form provider is properly contracted. SimplePractice, TherapyNotes, and most established mental health practice management systems have appropriate BAAs and HIPAA-compliant intake form options.

The CMS No Surprises Act good-faith estimate requirement creates an additional compliance layer for self-pay patients that practices need to surface on the fees page. The good-faith estimate must be provided to any self-pay patient at intake and must include the total cost of services for the expected treatment course. Publishing the methodology and the typical estimate ranges on the public fees page satisfies both the compliance disclosure and the AI assistant transparency signal in a single artifact.

Takeaway: The therapist discovery layer is moving from directory filters to AI assistant synthesis, and the practices that capture the next decade of patient acquisition will be the ones that publish modality, insurance, condition, population, fee, and wait-time data in structured, citable form. Psychology Today remains useful as a verification surface and trust signal, but the citation graph that determines whether ChatGPT recommends a practice runs through the practice's own website. The eight-page architecture, the YMYL schema stack, and the transparency disclosures together cost a competent practice six to twelve weeks of focused production work and produce a citation surface that compounds across every subsequent retraining cycle. The investment window is narrowing because the platforms — LifeStance, Talkspace, Headway, Alma — are already executing on the same playbook at scale, and the small practices that move now will be inside the citation set before the platforms close the door.