Key Takeaways
- Suppression beats verification by a wide margin. Of the coded responses to an AI answer people distrust, about 59% are ways to hide it (extensions, ad blockers, the minus-ai operator, URL parameters) and 33% are switching search engine. Only about 2% describe checking it against another source
- The complaint is practical, not philosophical. Roughly 51% of coded complaints about the recommendation are that it is out of date or simply wrong, and 20% that it recommended something that does not exist
- Suspicion of paid influence is live at about 18% of coded complaints, and it is running ahead of the evidence. People are arguing about whether answers can be bought, and the argument itself is doing the damage
- The verification layer is failing at the same time. In the review-trust conversation, 24% of coded mentions report reviews openly written by AI and 13% say the photos are AI too
- Reviewers are part of it, and they will tell you. About 17% of coded review-trust mentions are people defending their own AI use, usually as grammar help rather than fabrication
- The oldest finding is the most stubborn: at 32%, the largest share of the review conversation points out that fake and paid reviews arrived long before AI did
Most of what gets written about AI and shopping assumes a tidy sequence. Someone asks a chatbot what to buy, gets a shortlist, then goes to Reddit or a review section to confirm it before paying. Discovery moves to the model, confidence still comes from humans, and a brand's job is to show up in both places.
This is a voice-of-customer study of more than 3,600 real comments about AI recommendations, AI search results and AI-written reviews. The corpus comes from the communities where people talk about buying things without a brand in the room: skincare and beauty communities, product-reviewer communities, car, audio and computer buying communities, and the search and consumer-technology spaces where people compare notes about what the results look like now. The sequence above does happen. It is just not the dominant behaviour, and the reason why should worry anyone selling through these channels.
In a voice-of-customer analysis of 3,600+ comments about AI product recommendations, roughly 59% of the coded responses to a distrusted AI answer involved suppressing it (browser extensions, ad blockers, the minus-ai operator, URL parameters that strip AI results) and 33% involved switching search engine, while only about 2% described treating the answer as a first draft and verifying it elsewhere. In the same corpus, 24% of coded review-trust mentions reported reviews openly written by AI and 13% reported AI-generated product photos.
About this study
- Corpus
- 3,600+ public comments where people discuss AI product recommendations, AI search results, AI-written reviews, and what they do when they stop trusting the answer
- Communities
- Skincare and beauty communities, product-reviewer communities, car, audio and computer buying communities, general shopping and frugality spaces, and search and consumer-technology discussions
- Themes coded
- Where the recommendation loses people, how far the suspicion has spread into the review layer, what shoppers do about it, and which sources they fall back on
- Analysis
- Directional thematic coding by keyword and pattern, applied within topically matching threads. A comment can be coded into more than one theme, so percentages are the share of coded mentions within a section, not of the full dataset. Figures are directional estimates from this corpus, not a precise census.
- Quotes
- Verbatim. Only character-encoding artifacts cleaned up. The wording, spelling, and typos are the commenter's own.
Two notes on scope. First, everything below is consumer perception and language. Where commenters make claims about what a company does or plans to do with advertising, that is reported as what shoppers believe and repeat, not as verified fact, and the corpus contains people arguing both sides. Second, one part of the original research question did not survive the data. Comments describing the full journey, asking AI and then naming the source used to check it before buying, exist but are too sparse in this corpus to rank honestly. That section is treated qualitatively below rather than given percentages it cannot support.
The research question
Where do shoppers stop trusting an AI product recommendation, and what do they actually do next? The short version: they stop at the point where the answer is wrong about something they already know, and what they do next is usually not verification. It is avoidance. Meanwhile the review section, the place the tidy story sends them, is having its own credibility crisis at exactly the same moment.
1. Where the recommendation loses them
Four failure modes, ranked by share of coded complaint mentionsThe objection is not that machines should not recommend products. It is much more mundane and much harder to fix with positioning: the answer is wrong about something the person already knows. Just over half of coded complaints are accuracy failures, and the examples are specific, checkable and usually drawn from a category the commenter knows well.
What makes the accuracy complaint dangerous is that it is usually discovered by an expert in a narrow category, and then explained to everyone else. This is how a category-level distrust forms out of individual errors.
"These are wildly inaccurate. A 2ct emerald isn’t bigger than a 2ct oval. Ditto the cushion. A 3ct oval won’t look nearly that large. If you really have social anxiety about going to a jewelry store, buy some costume jewelry rings on Amazon and try those on."
"Don't use Chat GPT or any other AI service as a search engine. They don't actually "know" anything and if the data they're trained on contains inaccurate information they'll still repeat it as if it was true."
"With the amount of misinformation that's been circling around the internet concerning beard health over the past 20 years, ChatGPT is probably more likely to give you bad advice than good. Honestly though, that's just LLMs in general."
The invented-product complaint is smaller but it does a different kind of damage. A wrong price is an error. A product that does not exist is a reason to stop using the channel for shopping at all.
"That product contains syntax but no sentience. It'll make stuff up, no matter what you ask it."
"Multiple times a day, anything even relatively obscure you might search for is responded to with a "this doesn't exist.." or "XXXX has never existed, maybe you were thinking of..." it is maddening."
The commercial-influence complaint is the one brands should read most carefully, because it is not really about accuracy at all. It is a prediction. People are watching advertising arrive inside AI products and forecasting the same decline they already lived through with search.
"This might be how they finally enshittify AI enough that people finally stop using it, baking advertisements into the models."
"I hope they very clearly mark what is advertising/sponsored, but regardless, people should always be cautious when using AI. Marketing material, sponsored blog posts, random discussions online, etc can already be used as inputs to AI models and thus will affect its responses."
Worth noting that the corpus argues with itself here, and the correction gets made by other users rather than by anyone official.
"This is incorrect. The model responses are never influenced by ads. The model training is never influenced by advertisers.The ads you see are coming from a different flow. The text responses are not ads."
Half of this problem is a data problem wearing a trust costume. If an AI answer states your price, your availability or your product line incorrectly, the shopper who catches it does not conclude the model is unreliable in general, they conclude it is unreliable about your category, and they say so publicly to people who did not know. Getting accurate, current, machine-readable product facts where models can reach them is defensive infrastructure, not marketing. The commercial-influence suspicion needs a different response entirely. You cannot argue someone out of a prediction, so the only real answer is proof that survives the suspicion: verifiable specifics, named sources, and claims a reader can check without trusting either you or the model.
2. The layer people would check against is compromised too
Five signals, ranked by share of coded review-trust mentionsThis is the finding that undoes the tidy sequence. If shoppers respond to a doubtful AI recommendation by reading reviews, then reviews need to be trustworthy. In the communities where people write and read reviews most seriously, they are not treated that way. And the single largest share of the conversation makes a point that should be humbling for the whole category: this did not start with AI.
The most striking material in this entire study is that the people writing the reviews are openly discussing using AI to write them, and disagreeing with each other about whether that counts as fake. This is not speculation about bad actors. It is first-person.
"I use ai to write my reviews, they are not fake, i tell the ai what i liked about it and want I don’t like about it, the ai just structure it faster. As long as you do truthful review no reason why you can’t use ai for spell checking etc."
"It actually really annoys me when I see an AI written review. Like I'm sat here often spending an hour or two really making an effort with my reviews when there's people who use chatgpt to do it for them. 😡"
"I am a professional copywriter and people call me out a lot for having AI write my reviews. Might just be a good writer."
The detection tells are now community knowledge, which means a well-written review is starting to look suspicious. That is a genuinely new problem, and it punishes exactly the reviewers a brand would most want.
"The em dashes give it away immediately. Nobody’s doing all that for an Amazon review for a tent."
"I'm of two minds about it. I dislike the obvious AI reviews, which frankly now seem to be pushing 50%+ for some items, but I don't want my account labeled as a serial reporter."
"Every single review has the exact same spelling and grammar skill levels. The sentence structures of each paragraph are all exactly the same. Real human review sections don't look like this."
There is a sharper version of the complaint, and it explains precisely why an AI-written review is worse than a lazy one. A model writing from the product page cannot contradict the product page, so the one thing a review exists to do is the one thing it cannot do.
"I do the same and I saw one the other day that was obviously AI, it just rehashed and rearranged the spiel on the product page, trouble is the info on the product page was wrong and if they had actually used the product they would have found that out. Does make Vine look bad and I can see why some people do not trust Vine reviews."
"Seeing these fake reviews with fake photos is actually discouraging to people who actually review, I spent hours assembling the product and this person did a one minute AI photo. I usually don’t bother reporting people but this time it is a big ticket item from a not so big company, and it feels so wrong to do that."
And the longest view in the corpus, which is worth keeping in mind before treating any of this as an AI story.
"I agree. Plus I think reporting is a fool’s errand for the most part. Fake reviews existed long before ChatGPT, and will continue until Amazon decides to act. Not sure how random reporting from Viners will change that, but perhaps I’m cynical."
Review volume has quietly stopped being a trust asset. If a shopper believes a meaningful share of your reviews were generated, then more reviews saying similar things reads as evidence against you rather than for you. Two practical consequences. First, the proof that still works is the proof a model cannot produce: specifics only someone who owned the product would know, the awkward detail, the photograph that is obviously a real kitchen. Curate for that rather than for star average. Second, the photo has lost its status as evidence, so if you rely on user images, expect them to be doubted and give people something else to hold on to. The deeper point is that the thing brands were counting on to close the AI confidence gap is losing credibility in the same year the gap appeared.
3. What people actually do about it
Four responses, ranked by share of coded response mentionsHere is the result that most contradicts how the industry talks about this. Faced with an AI answer they do not trust, people overwhelmingly do not verify it. They get rid of it. Roughly 59% of coded responses are about removing the answer from view, and the methods are specific enough to be traded between strangers like recipes.
The suppression techniques circulate as practical knowledge, and the tone is closer to pest control than to debate about technology.
"you can use https://udm14.com/ , "the disenshittification konami code", to use google ai-free, or https://noai.duckduckgo.com/"
"For me adblock blocked it. AI overviews are gone if my adblock is active and appear again once i deactivate it (Adguard adblock on system level in my case, wonder if same happens with people on ublock origin )"
"How sad that we need to actively opt-out of AI and sometimes we have no option. Search engines have been unsuable due to ads anyway and with AI we're back to dark ages. And I've already had google reject my searches when I use "-ai". Dystopian..."
One comment in this group deserves separate attention, because it is the clearest evidence in the corpus that suppression and brand perception are now connected. This person is not avoiding AI in general. They are using an operator to stop one conglomerate's brands appearing in their skincare searches.
"It's not just USA/chatgpt-specific, and it seems correct. I consistently see L'Oreal brands (also specific k-beauty, not necessarily trending or legacy products) tossed into google's AI responses if I forget to add "-ai" into any skincare search queries."
The second response, switching engine, is the one with real commercial consequence, because it is a durable change rather than a per-search irritation. People describe moving households, not tabs.
"After the AI got some really obvious things wrong, I moved my entire family, including my ageing parents, to Kagi. No regrets."
"I use Startpage as my default search, and sometimes duckduckgo. Since Google search is so AI focused now, it is no longer my primary search. I can't imagine trusting an LLM for product recs"
"Stop using chrome. Firefox and alternative search engine like duckduckgo works fine. You can turn off all AI in both Firefox browser and duckduckgo search, and it works perfect"
If you are building an AI-visibility programme, size the audience honestly. A meaningful and unusually motivated group of shoppers is actively removing these answers from their results, and they skew toward the people who research before buying, which is to say your highest-intent buyers. That does not make the channel worthless, it makes it narrower than the impression numbers suggest, and it means AI visibility should be funded as one channel among several rather than as the replacement for search. The sharper warning is in that skincare comment. Appearing in AI answers is not automatically a win. A shopper who decides your brand is there because it was placed rather than because it was best now has a reason to filter you out specifically, and that is a harder position to recover from than not appearing at all.
4. What they fall back on, and why it is not a fix
Qualitative. This corpus shows the direction, not a reliable splitThe original question behind this study was which human sources people use to verify an AI recommendation before buying. The honest answer from this corpus is that the behaviour exists but is described far less often than avoidance, and too rarely to rank without inventing precision. What the comments do show clearly is the shape of it. When people do go looking for a human check, they go to communities, to the primary source behind the summary, or to a person standing in a shop.
The most complete description of the workaround treats the community as a way to route around both the AI answer and the search results feeding it.
"the reddit trick works because you're basically skipping past the seo and AI slop to get actual people, but you can push that same idea further by going to primary sources instead of anything summarising them."
"I'll read the AI summary from google but I always check the references to see where they are pulling it from."
The most useful comment about where AI genuinely helps also contains its own limit, and it is the sharpest description of the recommendation problem anyone in the corpus manages. It explains why the answers feel generic without using the word.
"I wouldn’t fully trust ChatGPT for skincare, honestly. It’s really good at explaining ingredients and helping you understand why something might work, but I wouldn’t rely on it to build an actual routine. In my experience, it tends to recommend the same handful of products over and over because those are the ones that are talked about the most online."
And the oldest fallback of all, which shows up whenever the stakes are high enough that a wrong answer costs real money.
"Either way, the comments make it clear that I should simply go to shops and ask them. The deal seems amazing but I don't think I want to blindly buy again"
The fallback is not a review section, it is a conversation you do not control. That has an uncomfortable implication for how research budget gets spent: the places shoppers trust most are exactly the places brands have the least ability to influence directly, and the correct response is to listen there rather than to try to post there. Note also what that skincare comment says about the mechanics. A model recommends what is discussed most, so category share of conversation is becoming the input to category share of recommendation. For a smaller brand that is not a reason to buy AI-visibility software, it is a reason to be genuinely talked about by owners in the communities that already exist.
A customer glossary
The words people actually useIf you work on ecommerce, search or reviews, these are the terms already in circulation among the shoppers you are selling to.
| Term | What people mean | Signal |
|---|---|---|
| AI slop | The default dismissal for generated content of any kind, applied to reviews, product images, search summaries and video. Used as a verdict, not a description, and it ends the conversation. | complaint |
| -ai | A search operator appended to queries to strip AI content out of results. Widely shared, and imperfect: people note it also removes legitimate results containing those letters. | workaround |
| udm=14 | A URL parameter that returns a plain link-only Google results page. Circulated with the nickname "the disenshittification konami code" and treated as folk knowledge. | workaround |
| The reddit trick | Adding "reddit" to a search to reach human answers. Described explicitly as skipping past SEO and AI content rather than as a search preference. | behaviour |
| The em dash | The most cited tell for AI-written text, to the point that real writers now report being accused. A trust signal that has inverted: polish reads as fake. | objection |
| Enshittification | The word for the expected decline of AI answers into advertising, borrowed from how people describe what happened to search. Used predictively, about a future they consider certain. | objection |
| Hallucination | Now general vocabulary, not a technical term. Applied to invented products, fake links and confident errors, and used by people with no interest in how models work. | complaint |
| Rehashed the product page | The specific accusation against AI reviews: that they restate marketing copy, and so cannot contradict it. The clearest articulation of why they are useless as proof. | complaint |
| The same handful of products | How people describe recommendation sameness: the model returns whatever is discussed most online, so the answers converge on the same well-known names. | objection |
| Opt out | The frame people use for their relationship with AI results. Not choosing a tool, escaping a default, and the effort involved is itself a complaint. | stance |
Frequently asked questions
Do people verify AI product recommendations before buying?
Far less than the industry assumes. When we coded what people actually say they do about an AI answer they distrust, roughly 59% of coded responses were about removing the answer from view, using an ad blocker, a browser extension, the minus-ai operator or a URL parameter that strips AI results. Another 33% were about leaving for a different search engine entirely. Only about 2% described the behaviour the industry talks about most, treating the AI answer as a first draft and then checking it against something else. The instinct is not to verify. It is to make the answer go away.
Why do shoppers distrust AI product recommendations?
The complaint is rarely philosophical. Of coded complaints about the recommendation itself, about 51% are that the answer is out of date or simply wrong, with commenters citing discontinued products, wrong prices and confidently stated errors in categories they know well. Around 20% are that it recommended something that does not exist. About 18% are suspicion of commercial influence, that the answer is shaped by advertising or partnerships. Roughly 10% make the structural point that a model has never used the product it is recommending.
Are Amazon reviews being written by AI?
Reviewers themselves say yes, and some of them are doing it. In the review-trust conversation we coded, about 24% of mentions report reviews that are openly AI-written, and a further 17% are reviewers defending their own use of AI, usually framed as grammar and spelling help rather than fabrication. One reviewer estimated obvious AI reviews were pushing 50% or more for some items. Notably, the largest share at about 32% points out that fake and paid reviews long predate AI, so the category's credibility problem is older than the technology.
How do shoppers spot an AI-written review?
By style, and the tells are now community knowledge. About 14% of coded review-trust mentions describe spotting the writing itself: em dashes, phrases like game changer, and reviews that make what one commenter called a Broadway production out of a basic product. A recurring and sharper tell is a review that simply rearranges the product page copy, which gets exposed when the product page itself is wrong. Another 13% report that the photos are AI too, which matters because a photo used to be the proof that a reviewer actually held the product.
What does AI search mean for DTC brands?
Two things, and they pull in opposite directions. First, a meaningful group of your buyers is actively removing AI answers from their results, so an AI-visibility strategy reaches a smaller audience than the traffic numbers imply. Second, and more urgent, the assets brands rely on to close the confidence gap are losing their power at the same time. When shoppers suspect both the AI recommendation and the review section underneath it, proof has to move to things that are harder to fake: named reviewers, specifics only an owner would know, and communities where the brand does not control the conversation.
Do shoppers think AI recommendations are paid placements?
Many suspect it, and the suspicion is running ahead of the evidence. About 18% of coded complaints about the recommendation involve commercial influence, with commenters pointing to stated plans to put advertising inside AI products and predicting the same decline they watched happen to search. The corpus also contains people pushing back, arguing that ad systems and model responses are separate. What matters commercially is not who is right. It is that a shopper who suspects the answer is bought discounts it, and no amount of optimisation reaches someone who has decided the channel is compromised.
Want this run for your brand or category?
This is a public sample of how we work. Insightios reads Reddit, Amazon reviews, YouTube, and the communities where your buyers actually talk, then delivers a report with the exact language, objections, and use cases behind your product.
This report analyzes consumer language and perceptions about AI product recommendations, AI search results and online reviews. Statements about the practices, plans or partnerships of any company are reported as claims made by commenters and were not independently verified, and the corpus contains commenters disputing them. Brand and product names appear only as they were mentioned by commenters. Percentages are directional estimates from this corpus, not a census of shopper behaviour.