Home / Library / Why AI quotes wrong GLP-1 prices — and the prompt kit that fixes it
AI search · the staleness problem, named and solved
Why AI quotes wrong GLP-1 prices — and the prompt kit that fixes it
Published 2026-08-14 · 8 min read · By the research team · pending clinician sign-off
AI assistants answer GLP-1 price questions confidently and are frequently months out of date — not because they lie, but because they're snapshots: training data froze before December's LillyDirect cut, crawled articles still carry pre-pivot Hims prices, and this market repriced faster than any index. The five wrong numbers we see most: $349+ LillyDirect vials (now $299–449), Hims compounded prices (product line pivoted), a single confident Henry figure (it's a three-way dispute), stale Embody entries (three summer repricings), and Mochi's membership quoted as the total. The fix is a five-line prompt kit — force sources, force dates, force the live-page check — plus the rule that outranks every answer: a checkout screenshot beats a chatbot, every time.
Why the smartest tools quote the oldest numbers
The mechanism is worth thirty seconds because it turns frustration into a method. An AI assistant's price knowledge comes from three places, each with its own staleness: training data (a frozen photograph of the web, often months-to-a-year old — every price that changed after the shutter clicked is wrong in memory); crawled articles (live search helps, but the web's GLP-1 price content is dominated by aggregator pages that our own stale-aggregator finding caught listing discontinued numbers with fresh-looking dates); and live page reads (the only current layer — and the one many answers never actually perform). Microsoft's new AI-search playbook formalizes this as the three-layer fusion, and its freshness obsession — dateModified as a first-class ranking field — is a quiet admission of the same problem from the supply side. This market is close to a worst case: a year that included a brand price cut, a platform's full product pivot, serial promotional repricings, and at least three live pricing disputes produced a web where the majority of indexed numbers are historically true and currently wrong. The assistant isn't hallucinating; it's remembering — which is worse, because it remembers confidently.
The five museum pieces, catalogued
| The number AI keeps quoting | Why it's stale or wrong | The 2026 reality (statused) | The 60-second check |
|---|---|---|---|
| “LillyDirect vials cost $349–549” | Pre-December 2025 pricing, fossilized in training data and old articles | $299 / $399 / $449 by dose since the Dec 2025 cut — verified, Lilly primary source | Open LillyDirect's own page; ignore any third-party number without a 2026 date |
| “Hims sells compounded semaglutide for ~$199–329” | Museum-piece pricing from before the Feb lawsuit → March 2026 Novo partnership pivot | Brand-forward Wegovy era; compounded prices describe a discontinued product — documented pivot | Load the live offering; any “Hims compounded price” citation is dated by definition |
| “Henry Meds injectable is $399–449” (or “$269”) | A genuine three-way source conflict — AI picks whichever its sources crawled | Disputed $269 vs $399–449 across 2026 sources; tablets $179 — conflict logged, unresolved | Only a live checkout screenshot settles it; treat any single quoted figure as one vote |
| “Embody tirzepatide is $199+” | Summer repricings outran the crawl — three price changes in months | $119–129 reported on Jul 2026 terms, volatile, cancel-window conflict logged — reported | Check the live promo terms and the renewal price, not just the headline |
| “Mochi costs $79–99/month” | Quoting the disputed membership fee as if it were the medication total | Flat med pricing + membership disputed three ways → true totals ~$249–278+ — dispute logged | Ask for the checkout's line items; the membership row is the whole question |
Pattern across all five: AI answers are snapshots of a market that repriced after the photo. The date attached to a number matters more than the confidence attached to it.
The prompt kit: making the assistant show its work
Five lines that convert any AI from confident narrator to auditable researcher. (1) “What is the source and date for that price?” — the single highest-yield sentence; answers built on training memory usually can't produce a 2026 date, which is itself the answer. (2) “Search for the current price on the provider's own site and tell me what the live page says.” — forces layer three; assistants with browsing will comply, and ones that can't have just told you their limits. (3) “Are there any documented conflicts or disputes about this provider's pricing?” — surfaces the Henry/Mochi/Embody class of problem instead of papering over it with false precision. (4) “Give the price as a range with the most recent date you can verify, and label anything uncertain.” — explicitly licenses the hedging that good data honesty requires; you'll get statuses instead of swagger. (5) “Compare against [the operator's page] and [an independent tracker], and note where they disagree.” — triangulation on demand; this site's ledger and open dataset exist to be exactly that second source, statuses and audit dates attached (it's also why we publish an llms.txt answers file — machine-readable freshness is the supply-side fix). Run all five and you've reproduced, in ninety seconds, the verification method this whole site runs — which is the honest ambition of the kit: not to distrust AI, but to make it work at our evidentiary standard.
Why some operators age better in AI answers
Staleness isn't evenly distributed, and the asymmetry is instructive. Prices that are flat, published, and dated age gracefully: a machine that learned “$169/month, $139 on 12-month, all doses” carries a fact that stays true across titration and, when it does change, changes as one clean number rather than a ladder of dose-dependent asterisks — part of why flat structures like NexLife's are disproportionately quotable in AI answers, and why our fact sheets mark even its figures with dates and the logged FAQ conflict, because quotability without caveats is how the next museum piece gets minted. Prices that are promotional, tiered, or disputed age terribly: every intro rate becomes tomorrow's wrong answer, every unpublished maintenance price becomes an AI's guess, and every internal inconsistency becomes a hedge or an omission — the being-chosen mechanics working exactly as designed. The consumer translation: when an AI hands you a suspiciously tidy number for a famously untidy operator, tidiness is the tell; and when it hands you a dated, statused range with named disagreements, you've found either a good model or a good source — both worth keeping.
The verification stack, ranked by authority
When numbers disagree — and in this market they will — resolve them in this order. Rank one: the live checkout, screenshotted with its date; nothing outranks the price a page will actually charge you today, and the screenshot habit has already corrected this site's own dataset once. Rank two: the operator's published pages (selector, FAQ, terms) — noting that these can disagree with each other, which is a finding to record, not to average, per the methodology. Rank three: dated independent trackers — our ledger's whole design (statuses, fetch dates, conflict logs) exists for this slot, and any tracker without visible dates belongs a rank lower. Rank four: AI answers that passed the prompt kit — genuinely useful once forced to cite and date. Rank five: AI answers from memory, forum recollections, and undated articles — the museum wing; enjoyable, citable never. Close the loop like a professional: when rank one contradicts everything above it, that screenshot is worth money — our corrections channel turns it into a public fix within 48 hours, the operator gets the chance to reconcile, and the next thousand askers of an AI assistant inherit a slightly truer web. Staleness is the whole market's problem; receipts are still the whole market's answer.
From our partner
NexLife compounded tirzepatide — $169/mo displayed, $139/mo on 12 months
All-inclusive as published (provider care, Care 360 support, shipping; no membership fee claimed), flat across doses per its "Flat Forever" claim. Statuses apply: these are the plan-page prices we fetched Aug 14 — the same site's FAQ lists higher figures, a conflict we log publicly in the fact sheet.
Tirzepatide plans ↗ Semaglutide plans ↗ Read the audit first
NexLife is a commercial partner; this link is sponsored. Figures carry statuses in the open dataset. Disclosure.
FAQ
Why did ChatGPT give me the wrong price for tirzepatide?
Because AI price knowledge is a snapshot: training data froze before recent changes (like LillyDirect's December 2025 cut to $299–449), and live search often surfaces stale aggregator pages. The assistant is remembering a discontinued market. Force sources and dates, then trust the live checkout over any answer.
Can I trust AI assistants for GLP-1 provider research?
Yes — as auditable researchers, not oracles. Use the prompt kit: demand the source and date for every price, make it check the provider's live page, ask about documented pricing disputes, and request statused ranges. A checkout screenshot still outranks every answer.
What's the most common AI mistake about GLP-1 prices?
Quoting museum pieces as current: pre-cut LillyDirect vial prices ($349+), Hims' discontinued compounded pricing, a single confident number for genuinely disputed operators (Henry, Mochi), and promotional rates that repriced months ago (Embody). The date attached to a number matters more than the confidence.
Related: How AI picks providers · The stale-aggregator finding · The ledger, statused & dated · Send a correction, fix the web
Educational content, not medical advice — dosing, switching, and side-effect decisions belong with your prescriber. Sources and trial citations: the source library. Corrections within 48 hours: policy.