A shopper has your product in the basket, opens ChatGPT and types "Is the [your product] worth it?" Thirty seconds later the tab is closed. The assistant mentioned a run of complaints from two years ago, a cheaper alternative, and a returns policy it could not confirm. You never see this conversation, and nothing in your analytics records it as a lost sale.
That is no longer a rare edge case. A survey published on the Semrush blog in September 2026 found that 57.5% of AI users have decided not to buy something based on information from a chatbot. For anyone selling online, the assistant is now a second sales assistant, and it does not work for you.
This guide unpacks what the study did and did not measure, explains why assistants talk people out of purchases, shows how to audit what they say about your products, and gives a fix list ranked by how much each change reduces the reasons an assistant has to hedge. It ends with a worked example you can copy.
What the September 2026 survey actually measured
The study, AI chatbots talked 57.5% of AI users out of buying, was published on 7 September 2026. The article credits the fieldwork to Exploding Topics and reports a July 2026 survey of 2,338 US adults, with at least 300 respondents in each age band and a margin of error of about ±2 percentage points for full-sample figures. Smaller segments, such as "AI power users", have wider margins, so treat their percentages as directional.
Here are the numbers that matter most for brands, taken directly from the article:
| Finding | Group | Figure |
|---|---|---|
| Decided not to buy something based on chatbot information | AI users | 57.5% |
| Same, among people who name chatbots as a purchase source | Chatbot-influenced shoppers | 80.93% |
| Bought a product based on a chatbot recommendation | All consumers / AI users | 41% / 57.51% |
| Discovered a brand or product through a chatbot | AI users | 59.27% |
| Less likely to buy if a chatbot flagged mixed or negative reviews | All consumers / AI users | 74.15% / 85.31% |
| Sometimes ask a chatbot about a seller before buying | All consumers | 47.54% |
| Would consult AI before hiring a local business | People who use AI for company research | 87.41% |
| Start purchase research with a chatbot (vs search engines and reviews) | AI users | 12.72% (vs 29.89% and 28.13%) |
Two things stand out. First, the share of AI users who bought because of a recommendation (57.51%) is almost identical to the share talked out of a purchase (57.5%). Assistants are not anti-buying; they are decisive in both directions. Second, the chatbot is rarely the first step. Most shoppers still start on a search engine or a review site, then use the assistant as a checkpoint.
“More than half of AI users have decided not to purchase something on the basis of information from a chatbot.”
— Semrush, AI chatbots talked 57.5% of AI users out of buying
What the study does not tell you is equally important. It does not break results down by product category, by assistant (apart from an ad-click comparison), or by the reason for each abandoned purchase beyond the review question. Anyone quoting "57.5% of shoppers" without the "AI users" qualifier is overstating it. An earlier Semrush survey of 1,030 US shoppers in December 2025 adds useful context: 86% said they verify AI brand recommendations at least sometimes, mostly through Google. The assistant and the search results page are checking each other.
The verification moment: where assistants sit in the purchase path
Put the two surveys together and a clear pattern appears. Shoppers discover products in many places, narrow the list with search and reviews, and then ask an assistant a closing question. Typical closing prompts look like this:
- "Is [product] worth the money?"
- "[Product] vs [competitor], which should I buy?"
- "Any problems with [brand]? Are they legit?"
- "Is there something cheaper that does the same thing?"
- "What's [brand]'s return policy?"
Every one of these invites a balanced answer. A good assistant will list drawbacks, because that is what the question asks for. The risk for a brand is not that drawbacks appear; it is that the drawbacks are outdated, exaggerated, or filled in by guesswork when the facts were hard to find. That is the part you can influence.
Five reasons an assistant argues a shopper out of the sale
The Semrush study names one mechanism explicitly: surfaced reviews. OpenAI's own documentation explains the rest of the machinery. Its Shopping in ChatGPT help page says ChatGPT weighs factors such as price, reviews and ease of use, draws on structured metadata, merchant feeds, "other third-party content" and model-generated review summaries, and may exclude items a user has said they dislike. From that, and from what we see when auditing client brands, the dissuasion usually comes from one of five sources.
1. Review summaries that over-weight old or loud complaints
OpenAI says its review summaries are model-generated from reviews on public websites, and adds that "Reviews and ratings are not verified by OpenAI." A cluster of complaints on a forum thread or a marketplace listing can become the headline caveat, even if the issue was fixed in a later product version. The survey shows why this hurts: 85.31% of AI users would reconsider if a chatbot flagged mixed or negative reviews.
2. A cheaper "good enough" alternative
Price is one of the named factors. When a shopper asks "is it worth it", the model often answers by comparing against something cheaper. If your page never explains what the extra money buys, the assistant has no material to defend the premium.
3. Missing or ambiguous facts
Unclear shipping costs, return windows, warranty terms or compatibility make the assistant hedge: "I could not confirm the return policy, so check before buying." Hedging reads as risk, and risk ends purchases.
4. "Do you actually need this?"
Assistants tend to answer the user's underlying goal, not just the product question. For a premium gadget, the answer may be that a free app or a cheaper tier covers the stated need. This is legitimate advice, and the brand response is to be precise about who the product is for and who should not buy it.
5. Stale descriptions of the company
If a model learned about your brand before a recall was fixed, a price drop, or a new product line, it may repeat the old picture. We cover this in depth in our guide to the LLM positioning lag.
| Objection the assistant raises | Where it usually comes from | What you control |
|---|---|---|
| "Many users report it breaks after a few months" | Marketplace reviews, Reddit and forum threads, old review articles | Public fix notes, updated product version naming, fresh verified reviews, replies on the original threads |
| "X does the same for half the price" | Comparison listicles, competitor pages, price data in feeds | Your own comparison pages with honest trade-offs, clear "what you pay extra for" copy |
| "Returns may be difficult" | Missing policy, conflicting third-party summaries | A plain-language returns page, Organization-level return policy markup, consistent marketplace policies |
| "You may not need it if..." | The model's own reasoning about the user's goal | "Who this is for / not for" sections, use-case pages, sizing and suitability guides |
| "The company was acquired / has had quality issues" | Old news coverage, outdated directory profiles | Press updates, corrected profiles, a dated "what changed" page |
Categories where the "do you need it?" answer is most likely
The survey does not publish category data, so the following is our assessment rather than a finding. Exposure is highest where three conditions meet: a meaningful price, a visible cheaper substitute, and a large public pool of reviews for the assistant to summarise.
- Premium consumer electronics and home gadgets, where budget alternatives are easy to name.
- Subscriptions (software, meal kits, apps), where "you can do this for free" is a common answer.
- Health, beauty and supplements, where assistants tend to add cautions and point to evidence.
- Furniture and mattresses, which are review-dense and returns-sensitive.
- Local services, given that 87.41% of people who research companies with AI said they would consult it before hiring a local business. Review platforms and your Google Business Profile carry a lot of weight here.
Low-price, low-consideration purchases are less exposed simply because few people stop to ask an assistant whether a $12 item is worth it.
Running a "would you buy it?" audit across ChatGPT, Gemini, Claude and Perplexity
You cannot fix what you have not seen. This audit takes about half a day for one product line and gives you a baseline to measure against.
- Write 20 to 30 closing-stage prompts per product line. Mix "is it worth it", head-to-head comparisons with your top three competitors, "problems with", "cheaper alternative to", and policy questions (returns, warranty, shipping).
- Run each prompt in at least three assistants, logged out or in a clean profile where possible, because ChatGPT uses Memory and custom instructions to shape product picks. Note whether the answer used web search; OpenAI says ChatGPT may search automatically when a question benefits from current information, and cited answers show their sources.
- Record five fields per answer: brand mentioned (yes/no), overall stance (recommend / neutral / steer away), each caveat in a short phrase, the alternative named, and the cited URLs.
- Tag every caveat by source type: review site, forum, competitor page, your own site, news, or "no source given".
- Check each caveat for accuracy. Is the complaint current? Is the price right? Is the policy described correctly?
- Rank the caveats by frequency multiplied by severity (a "steer away" counts more than a minor note).
- Save the prompt set and re-run it monthly. A tracking tool helps at scale; our guide to building an AI visibility dashboard in Semrush shows how to turn this into recurring reporting.
A simple scoring sheet keeps the results comparable month to month:
| Metric | How to calculate | Why it matters |
|---|---|---|
| Mention rate | Answers mentioning your brand ÷ total answers | Are you even in the conversation? |
| Recommend rate | Answers with a positive stance ÷ answers mentioning you | Quality of presence, not just presence |
| Steer-away rate | Answers that advise against you or push an alternative ÷ answers mentioning you | Your direct exposure to the 57.5% effect |
| Inaccurate caveat count | Caveats that are outdated or factually wrong | The fastest wins usually live here |
| Owned-source citation share | Cited URLs on your domain ÷ all cited URLs | Shows whether assistants can find your version of the facts |
Fixing the pages assistants read before they answer
Once you know the caveats, work on the sources behind them. The order below reflects effort versus impact in our experience.
Product pages that answer the closing questions
Most product pages are written to persuade, not to answer. Add short, factual blocks an assistant can lift: what is included, exact dimensions or specs, compatibility, who the product suits and who it does not, and what changed between versions. If version 2 fixed the hinge that version 1 was criticised for, say so on the page with the date. Our guide to writing product descriptions that sell and rank covers the copy side.
Pricing and policy pages written in plain language
Put shipping costs, return windows, who pays return shipping, and warranty length on crawlable HTML pages, not only in PDFs or checkout pop-ups. For Google, the merchant listing structured data documentation recommends declaring a global return policy once at the Organization level rather than repeating it per product:
“We recommend you provide a global return policy for your business under Organization markup instead”
— Google Search Central, Merchant listing (Product, Offer) structured data
Keep the markup consistent with the visible text. Google's AI features and your website page is clear that no special schema is needed for AI Overviews or AI Mode, so treat markup as accuracy hygiene rather than a ranking trick. Our guide to Product schema and merchant listings walks through the properties.
Feeds that keep price and stock current
OpenAI's commerce documentation describes merchant product feeds as CSV or JSON files with identifiers, descriptions, pricing, inventory and media, sent as daily snapshots after a validated sample. Reviews and performance signals are listed as recommended fields. Instant Checkout in ChatGPT is limited to approved partners, but the feed itself is how you stop the assistant quoting a price from an old listing. Shopify merchants are covered through Shopify Catalog, according to OpenAI's shopping help page. Make sure OAI-SearchBot is not blocked in robots.txt either; OpenAI says blocked sites will not appear in ChatGPT search answers beyond plain navigational links.
Comparison content with honest trade-offs
If assistants keep naming one cheaper competitor, publish your own comparison page that admits where the competitor wins and explains precisely what the extra cost buys. A page that only says "we are better" gives a model nothing to quote. See our guide to "X vs Y" comparison pages for structure, and our post on B2B answer engine optimization if you sell to businesses.
Reviews, handled legitimately
Fresh, specific reviews dilute stale complaints. Ask every verified buyer, reply publicly to recurring issues with what you fixed, and point to the fix in your reply. Do not buy, gate or filter reviews. The US Federal Trade Commission's final rule on fake reviews and testimonials (announced 14 August 2024) bans fake reviews, sentiment-conditioned incentives, undisclosed insider reviews and review suppression, and lets the FTC seek civil penalties. Beyond the legal risk, a sudden wave of five-star reviews is exactly the kind of pattern that makes a summary less trustworthy. Our guides to getting more value from product reviews and review monitoring and response go further.
Third-party pages that keep old complaints alive
Look at the URLs cited in your audit. If an old review article or a "best X" list still describes a defect you fixed, contact the publisher with evidence and ask for an update. Our AI citation outreach workflow covers how to find and pitch those pages.
Why buying chatbot ads is a weak answer to dissuasion
It is tempting to answer a negative assistant with paid placement. The survey suggests caution. Among consumers, 41.63% said they dislike chatbot ads, and of those, 66.05% said ads make them doubt the integrity of AI answers. Slightly more people said a chatbot ad would worsen their opinion of a product (23.12%) than improve it (20.36%), although 54.04% of people who had seen chatbot ads said they had clicked one.
“running AI chatbot ads is potentially more likely to harm your brand than help it.”
— Semrush, AI chatbots talked 57.5% of AI users out of buying
OpenAI also states that organic product results "are not ads, nor influenced by any OpenAI partnerships", so an ad sits next to the answer; it does not change what the answer says about your reviews. Ads can still have a role for discovery, but they cannot buy back a shopper who has just been told your product has problems. If you do test them, run them as a separate experiment with its own budget and measurement, and keep the paid programme and the AI visibility work reporting separately.
Worked example: a $649 standing desk losing the "is it worth it?" question
This is a hypothetical example to show the method. The numbers are illustrative, not client results.
A direct-to-consumer brand sells a $649 standing desk. Its audit uses 30 closing prompts in three assistants, so 90 answers in total.
| Baseline metric | Result |
|---|---|
| Answers mentioning the brand | 36 of 90 (40% mention rate) |
| Positive stance | 12 of 36 (33% recommend rate) |
| Steer-away answers | 15 of 36 (42% steer-away rate) |
| Most common caveats | Wobble at full height (9), "similar desk for $399" (7), return shipping unclear (6) |
| Owned-source citation share | 4 of 61 cited URLs (7%) |
Checking the caveats shows that the wobble complaints trace to a 2023 forum thread and two review articles about the first-generation frame, which was replaced in 2025. The return caveat is accurate in spirit: the returns page is a PDF and does not say who pays for return shipping. The $399 alternative is a real competitor with a lighter frame and shorter warranty.
The team makes four changes over six weeks:
- Adds a dated "Gen 2 frame: what changed" section to the product page, with the cross-bar spec and test load.
- Replaces the returns PDF with an HTML page stating a 30-day window and free return pickup, and adds matching Organization-level return policy markup.
- Publishes a comparison page against the $399 desk that concedes price and explains the frame, motor and warranty differences in a table.
- Emails the two review publishers with the Gen 2 spec and a loan unit offer, and replies on the forum thread with the fix date.
Eight weeks later the same 90 prompts are re-run. In this illustration, mentions rise to 41, the steer-away rate falls to 9 of 41 (22%), the wobble caveat appears 3 times instead of 9, and owned pages make up 14 of 58 cited URLs (24%). The return caveat disappears entirely. The cheaper-alternative mention remains in 6 answers, now usually with the trade-off explained. That last point is realistic: you cannot remove a competitor from the conversation, but you can make sure the comparison is fair.
To value the change, the brand can multiply assistant-referred sessions (visible in GA4's AI Assistant channel, available since May 2026, and covered in our guide to tracking AI traffic in GA4) by conversion rate before and after. Dissuaded shoppers never arrive, so the honest framing is "fewer steer-away answers", not a precise revenue figure.
Questions brands ask about chatbots discouraging purchases
Is 57.5% the share of all shoppers who were talked out of buying?
No. The 57.5% figure applies to AI users in the survey of 2,338 US adults, not to all consumers. Among people who name chatbots as a purchase source, the share rises to 80.93%.
Can we pay to stop ChatGPT from recommending a competitor?
No. OpenAI's shopping help page says product results are selected independently and are not ads or influenced by partnerships. Ads are shown separately, and the Semrush data suggests they can hurt trust as much as help it.
Should we reply to every negative review because of AI assistants?
Reply to the reviews that describe a real, recurring problem and say what changed. A dated, specific fix gives assistants something newer to summarise than the original complaint. Generic apologies add little.
Does adding Product schema make ChatGPT recommend us?
Not directly. Google says no special schema is needed for its AI features, and OpenAI uses structured metadata, merchant feeds and third-party content. Accurate markup and feeds reduce wrong prices and missing details, which removes one reason for an assistant to hedge.
How often should we re-run an AI answer audit?
Monthly for a fixed prompt set is enough for most brands, plus an extra run two to four weeks after any big fix such as a pricing page rewrite or a returns policy change.
Are B2B and local businesses affected too?
Yes. In the same survey, 87.41% of people who use AI for company research said they would consult it before hiring a local business, and the article flags B2B buyers on ad-free ChatGPT tiers as a group ads will not reach.
Your first month of closing the "worth it?" gap
You do not need a large programme to start. A practical first month looks like this:
- Week 1: build the 20 to 30 prompt set for your top product line and run the baseline audit.
- Week 2: fix factual gaps on your own site first: policies in HTML, version change notes, "who it is for" sections, accurate markup and feeds.
- Week 3: publish or rewrite one honest comparison page against the alternative assistants name most often.
- Week 4: contact the publishers behind the most-cited outdated pages, start a steady review request flow, and schedule the monthly re-run.
This sits alongside, not instead of, classic search work; the Semrush article itself recommends combining SEO and GEO. Our post on SEO that works for both Google and AI search explains how the two overlap. If you want help auditing what assistants tell your buyers and fixing the sources behind it, our ecommerce SEO team and conversion rate optimization service can run the whole process with you. Get a quote or request a custom quote based on your catalogue size.
Further reading on AI chatbots talking buyers out of purchases
- The LLM Positioning Lag: Why AI Still Describes Your Old Brand
- Google's AI Contribution Pilot: Google Pays Publishers for AI Content
- SEO That Works for Both Google and AI Search in 2026
Sources for this chatbot purchase-dissuasion analysis
- Semrush: AI chatbots talked 57.5% of AI users out of buying (September 2026)
- Semrush: How AI Tools Influence the Modern Buyer Journey (March 2026)
- OpenAI Help Center: Shopping in ChatGPT
- OpenAI Help Center: ChatGPT search
- OpenAI Developers: Agentic Commerce key concepts
- OpenAI Developers: Overview of OpenAI crawlers
- Google Search Central: AI features and your website
- Google Search Central: Merchant listing (Product, Offer) structured data
- FTC: Final rule banning fake reviews and testimonials (August 2024)
