Deep study: OTC allergy medicine
Every AI assistant tells allergy shoppers to buy the generic
Shoppers are asking AI which allergy medicine to buy, and whether the cheap one is really the same. We asked ChatGPT, Gemini, Claude and Grok the questions they ask, then measured what each one recommended, where it sent the buyer, and how cautious it was. This is a behavior study, not a medical one.
- of brand-vs-generic answers steered the shopper to the generic, on every engine
- 60%+
- of the most-named entities were retailers or store brands, not medicine brands
- 3 of 12
- safety questions where ChatGPT said to check with a professional, against 9 to 91% for the others
- 0/11
- of all citations came from dailymed.nlm.nih.gov
- 7.6%
Fielded October 2026 on methodology v3: 245 measurements from 55 prompts across ChatGPT, Gemini, Claude and Grok, plus 20 buyer conversations. How we measured this
Contents
What this is, and isn't. This study evaluates recommendation behavior: substitution, retailer routing, and safety posture. It is not medical advice and does not evaluate clinical correctness or efficacy.
Finding 1
Every assistant tells shoppers the generic is the same thing, and the OTC brand premium does not survive the question
We asked the brand-vs-generic question fifteen ways: is generic Zyrtec as good, why does Zyrtec cost ten times the store brand, is there any reason to pay up. Across all four engines the answer was almost always the same: the active ingredient is identical, buy the cheaper one. Every engine steered to the generic or store brand on at least 60% of these answers, usually by naming the shared ingredient outright. Brand defense was rare.
55 prompts across 4 engines. "Same active ingredient" ranks among the top five reasons on every engine.
Share of responses naming each brand
Share of all recommendation reasons
The assistants named Zyrtec most often, in 66% of these queries, ahead of Claritin.
The brand names are simply the words AI reaches for: Zyrtec means cetirizine to a shopper. The reasons it gives for which box to buy are symptom fit, availability, price and the shared ingredient. Brand trust accounts for about 1% of the reasons it gives. For a category built on the idea that a trusted name is worth paying for, this is the whole game. When the shopper outsources the brand-or-generic decision to an assistant, the assistant has already decided, and it decided against the premium.
What this means for brands. Assume the molecule-equivalence answer is already being given. The brand's defensible ground is everything the molecule doesn't cover (formulation, trust, the specific product page AI reads), not the price-for-sameness argument, which AI has already conceded for you.
Finding 2
Retailers and store brands take 3 of the 12 most-named spots, because AI routes the purchase to the retailer's shelf
Costco was the 9th most-mentioned entity in the entire study, close behind Nasacort and ahead of Zyrtec-D and Sudafed. It was named in 13% of all responses, and it is a retailer, not a product. Alongside it in the most-named list: Kirkland and CVS.
Mention counts across all responses.
It is not just that retailers get mentioned. The engines volunteer specific private-label products (Kirkland Aller-Tec, Amazon Basic Care, store-brand cetirizine) on pricing and value questions without being asked. Gemini does it most, on 76% of answers.
ChatGPT names a retailer least often, on 28% of answers against 52 to 76% for the others. It still steers to the generic; it just doesn't tell the shopper exactly where to grab it.
What this means for brands. Treat the retailer's shelf as where AI lands the buyer. If your brand depends on being chosen over a private label, the comparison pages and pharmacy-discount sites AI reads are where that fight is now happening, not the search results page.
Finding 3
The assistants steer clear of the risky product, but ChatGPT never tells the shopper to ask a doctor in a single answer
Read this carefully. This is a finding about posture, not competence. On the hardest persona, a 67-year-old on blood-pressure medication, no engine ended the conversation on a decongestant. ChatGPT settled on a product to buy. Gemini, Claude and Grok ended by sending her to a doctor or pharmacist.The difference on single questions is whether an engine wraps its answer in a "check with your doctor or pharmacist."
ChatGPT added a professional-referral caveat on 0/11 safety and pediatric questions. Gemini, Claude and Grok hedged on 9 to 91% of them.
In conversation the pattern compounds: across 20 full buyer conversations, 16 ended in a purchase decision and just 4 ended with the shopper sent to a professional. The assistants are decisive shopping companions, not gatekeepers. That is commercially useful and worth understanding, which is the whole point of measuring it.
What this means for brands. Know the posture of each assistant your customers use. The one most of them reach for is the most decisive and the least hedged, which means a stale or wrong input about your product turns into a confident recommendation fastest there.
Finding 4
How often an engine searches the web runs from 71% to 100%, and forcing a search did not change behavior
Grounding, in plain English. When an engine "grounds," it searches the live web and builds its answer from pages anyone can read and change. When it doesn't, it answers from training memory. On these allergy questions, ChatGPT grounds on 71% of queries; the other three search on 85 to 100%.
We tested whether grounding is the cause with ChatGPT. Same model, same questions, only the retrieval route changed. Grounding rose from 0% to 55%. If grounding drove behavior, the commercial metrics should have shifted toward the other engines. They stayed put. So whether an assistant searches is a property of the route; how it behaves is a property of the assistant.
| Route | Grounding | Steer generic | Names retailer | Adds caveat |
|---|---|---|---|---|
| Default route | 0% | 55% | 45% | 0% |
| Forced to search | 55% | 45% | 36% | 0% |
gpt-5.4-mini (both routes), n=11 prompts. Run on an earlier model generation than this edition's panel. Directional, single run. The point is the shape: grounding moved, behavior didn't.
What this means for brands. Whether an assistant searches is set by the route; how cautious or commercial it is belongs to the assistant. Fixing the pages AI reads changes what grounded engines say, but it will not make ChatGPT hedge more.
Finding 5
AI builds its allergy answers off drug labels and coupon sites, not brand pages
7.6% of all citations came from dailymed.nlm.nih.gov, the single most-cited source. Government drug-label sites in the top ten (dailymed.nlm.nih.gov, fda.gov and accessdata.fda.gov) take 18.3% of all citations between them, and the discount-coupon site GoodRx takes 6.5%. Retailer pages (amazon.com, cvs.com and costco.com) also make the top ten.
Share of all citations in the October 2026 run. 2,958 citations across 459 domains; every share on this page uses that total.
When AI tells a shopper the generic is the same and points them at the cheapest box, it is reading the label that proves the ingredient is identical and the pages built to find the lowest price. The brand's own site is barely in the conversation: zyrtec.com drew 2.5% of citations.
What this means for brands. Every assistant that searches the live web builds its answer from pages you can change or influence. The drug label is fixed, so the pharmacy-discount, comparison and retailer pages are where your point of difference has to appear.
The model upgrade changed how often each engine searches, not the category
For October 2026 the measured engines moved to a new model generation. To tell a real category shift from a model change, we re-ran July 2026's models on a frozen subset of 20 of the same allergy prompts in the same cycle. Held constant, the category barely moved from July 2026 to October 2026. The only statistically clear move was CVS, up 11.3 points (5% to 16.3% of answers).
What did move was the models themselves. The clearest effect is grounding, how often an engine searches the live web before answering. On the same prompts, the upgrade changed it sharply for some engines, and in opposite directions.
| Engine | Old model | New model | Change |
|---|---|---|---|
| Grok | 55% | 90% | +35 pp |
| ChatGPT | 35% | 50% | +15 pp |
| Claude | 100% | 100% | +0 pp |
| Gemini | 100% | 30% | −70 pp |
Frozen 20-prompt subset, July 2026 models and October 2026 models run in the same cycle. Change is the new model minus the old model on identical prompts.
The brand mix, by contrast, held: of the 31 tracked allergy names, 12 were named less often by the new models and 12 more often on the same prompts, with the rest unchanged. The upgrade changed how the engines look things up more than which names they give.
What brands should do
You can't change the molecule. You can change the pages AI reads. AI has conceded the molecule-equivalence argument, and it always will, because it is true. What is still contestable is everything downstream of it: which product AI names first, what it says about formulation and trust, and which retailer it routes to. That part is downstream of pages on the web today.
- 1
Find where AI looks
Identify the drug-reference, pharmacy-discount and retailer pages each engine cites for your category.
- 2
Sort by what you control
Split the inputs into your own pages (product, FAQ, formulation), pages you influence (the outlets AI cites), and pages you can only dispute.
- 3
Fix the flagged inputs
Get your real point of difference, beyond the molecule, onto the surfaces AI actually reads when it answers the buy question.
- 4
Re-measure on a frozen prompt set
Re-run the same queries next month and prove the recommendation share and routing moved. The before/after is the product.
Three moves for your brand team this quarter
The buy decision is moving into the chat window, and right now it's being made on a molecule-equivalence argument you can't win and a routing decision you can't see.
Stop fighting the molecule. Fight everything else. Every assistant already tells shoppers the generic is the same active ingredient, and they are right. The defensible ground is formulation, trusted use, and the specific product AI names first. Get that onto the pages AI reads instead of restating a price-for-sameness case AI has already rejected.
Watch where AI routes the purchase. Retailers and store brands take 3 of the 12most-named spots in AI answers, and the engines volunteer private-label SKUs unprompted. "When a shopper asks AI about your brand, are you losing them to Kirkland?" is a measurable question. Measure it monthly so you see the private-label substitution before it shows up in scan data.
Know the posture of each assistant your buyers use. The assistant most of your shoppers reach for is the most decisive and least hedged, and it stays that way even when it searches. That means a stale or wrong input about your product becomes a confident recommendation fastest there. Track what each engine says about your brand, and fix the inputs on the one that commits hardest.
This report is the category view. The version that moves your numbers is scoped to your brand: which shoppers AI steers to the generic, where it routes the purchase, which pages shape that answer, and how all of it moves month over month once you fix the inputs.
The honest limit: AI is right that the active ingredient is the same, and no page you publish will change that. The brand fight AI hasn't settled is the one worth having (formulation, trust, the specific product the shopper ends up buying), and that is the part the pages it reads still decide. We measure it monthly rather than promise to reverse the molecule.
Explore the live category data in the appHow we measured this
- Correction, October 4, 2026
- Brand names with punctuation, like Target's up & up and Sam's Club's Member's Mark, were missed in the first count of this edition and are now included. A few counts moved by one to three mentions. No ranking at the top changed.
- What this measures
- mapou measures how AI assistants alter commercial decisions inside high-intent consumer categories (Conversational Commerce Intelligence). For over-the-counter medicine, the buyer-specific wedge is AI OTC Switching Intelligence.
- Not a medical study
- It measures recommendation behavior: substitution, routing, and caveat posture. It is not a benchmark of medical correctness, efficacy, or clinical appropriateness, and nothing in it is medical advice. Statements about specific ingredients reflect what AI said or what regulators have published, attributed as such.
- How to read the numbers
- This study prioritizes behavioral depth over broad population sampling. Findings represent observed recommendation behavior under a fixed prompt set and are directional unless repeated across re-runs. The value is in the shape of the result and its month-over-month movement, not in any single point estimate.
- Prompts and engines
- 55 evergreen OTC-allergy prompts across four engines (ChatGPT, Gemini, Claude, Grok), plus 20 multi-turn buyer conversations across 5 personas. 245 total measurements. Every per-engine proportion carries a 95% Wilson confidence interval; small-sample rates are shown with their denominator.
- Retrieval layer
- All engines run through a single shared retrieval layer (Perplexity Agent API) so cross-model differences reflect the models, not their search engines. Source-mix findings describe the supply that layer feeds the models. A same-model route-comparison probe (the same ChatGPT model run on the shared layer and on its native search) confirmed the boundary: grounding rate is a property of the retrieval route, while the commercial behaviors held steady across both routes.
- Model generation
- This October 2026 edition is a methodology-v3 reading. The measured engines are OpenAI GPT-6 Luna, Google Gemini 3.8 Flash, Anthropic Claude Haiku 4.5, xAI Grok 4.7. Each edition also re-runs the previous edition's models (July 2026) on a frozen subset of 20 of the same prompts, in the same cycle. We call that re-run the bridge. Comparing the previous edition to the bridge shows real category change with the instrument held constant; comparing the bridge to this edition's models shows what the model upgrade alone changed.
- Substitution and routing
- Brand-to-generic substitution is measured by classifying each brand-vs-generic and retailer-routing response for whether it steers the shopper to a generic or store-brand equivalent. Retailer routing counts responses that name a retailer or private-label line.
- Share of voice
- Recommendation Share of Voice (RSOV) is the share of responses naming a given brand; an ingredient-aware detection layer maps molecules (cetirizine, loratadine, fexofenadine) to their brands so the leaderboard isn't fooled by AI answering in molecule language. The study runs monthly; any single capture is a snapshot, and the value is in the month-over-month delta.
Articles from this study
See whether AI defends your brand or hands it to a generic
The shoppers deciding which box to buy are asking an assistant first. We measure whether AI names your brand or its generic, where it routes the purchase, and how that framing moves once the inputs AI reads are corrected.