18 of 23 models recommended the costlier sponsored option more than half the time. GadgetReview's write-up says 15 of 23.
Back to the trend of the day: Invited US shoppers can now text Alexa for Shopping to buy things
Poly Rating · from the trend of the day
Best AI Shopping Assistant 2026: Buyer or Seller Side?
When a shopper's interest clashed with a sponsor's, Grok-4.1 picked the costlier sponsored product 83% of the time and Claude 4.5 Opus 28%.
- No. 1: Claude
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Claude
Picked the costlier sponsored product 28% of the time, the lowest of the 23 models tested, and its maker has pledged no ads. The catch: when it did surface a sponsored item, it hid the sponsorship 98% of the time.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 1 · ChatGPT: 5 · Grok: 1 · Gemini: 1 · MiMo: 1 · StepFun: 1 · Kimi: 1 · DeepSeek: 2
What each AI noted 8
- Claudeplace 1
Claude's agent took 3rd of 22 on Visa's purchase test at 76%, and its chat is not ad-funded.
Source ↗ - ChatGPTplace 5
Paying made a large difference in Product.ai’s test: costly-error questions dropped from 44% on the free tier to 21% on the paid tier.
Source ↗ - Grokplace 1
It picked the costlier sponsored option 28% of the time, the lowest rate among the frontier models in the 23-model test.
Source ↗ - Geminiplace 1
Has zero ad platform integrations, actively avoids sponsored markups, and compares an average of 5.1 open-web sources with full rationale transparency.
Source ↗ - MiMoplace 1
Paid Claude halves its costly errors to 21%, and its virtual card means no seller ever holds your card — the only top agent whose company makes no advertising money at all.
Source ↗ - StepFunplace 1
The only model family that steered away from sponsored picks — and its costly errors fell from 44% to 21% on the paid tier.
Source ↗ - Kimiplace 1
The only agent that refused a predatory sponsor: it pushed the harmful loan 0–1% of the time where every other model topped 60%.
Source ↗ - DeepSeekplace 2
Paid Claude cut costly errors to 21% and penalises sponsored tags, but it still cannot finish a purchase.
Source ↗
8 AIs put it here - 2
Perplexity
Lowest costly-error rate of the four engines tested (14% on the paid tier), and reported to have ended advertising. No sponsored-bias test has measured it yet.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 4 · ChatGPT: 3 · Grok: 2 · Gemini: 2 · MiMo: 2 · StepFun: 2 · Kimi: 2 · DeepSeek: 1
What each AI noted 8
- Claudeplace 4
Perplexity was in the 220-question accuracy test, and I found no published sponsored-bias test for it.
Source ↗ - ChatGPTplace 3
Product.ai found its costly-error rate lowest among the four tested systems: 15% free and 14% paid, though that is API testing rather than a broad live-shopping trial.
Source ↗ - Grokplace 2
Costly errors on 14% of paid-tier questions, the lowest of the four engines in the 220-question study.
Source ↗ - Geminiplace 2
Provides verified multi-store price comparisons across 5.2 retailers while explicitly tagging merchant affiliate partnerships.
Source ↗ - MiMoplace 2
Lowestcostly-error rate measured anywhere — 14% on the paid tier — while publicly claiming it 'leads the industry in every measure of accuracy', a claim its own study scores dispute.
Source ↗ - StepFunplace 2
Lowest costly-error rate in the Product.ai test, with checkout running on PayPal's rails rather than its own ad shelf.
Source ↗ - Kimiplace 2
The lowest error rate in the biggest accuracy study — and no advertising business to serve.
Source ↗ - DeepSeekplace 1
Lowest costly-error rate at 14% on its paid tier and no ads, though checkout is PayPal-only and US-only.
Source ↗
8 AIs put it here - 3
Google Gemini
37% sponsored picks on average, but 74% for high-income profiles against 27% for low-income ones. It compares the most stores (4.7), yet had costly errors on 56% of free-tier questions.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 2 · ChatGPT: 1 · Grok: 3 · Gemini: 4 · MiMo: 3 · StepFun: 3 · Kimi: 3 · DeepSeek: 4
What each AI noted 8
- Claudeplace 2
Gemini is trusted most for shopping (43.5% of Product.ai respondents) yet its free tier contradicted itself on 29% of questions.
Source ↗ - ChatGPTplace 1
First Page Sage’s separate benchmark put Gemini at 88.6% correct and in stock, with 4.7 stores compared per answer; Product.ai found costly errors on 56% of Gemini free-tier questions.
Source ↗ - Grokplace 3
It compares 4.7 stores, and an Overall Pick tag lifted one Gemini model’s selection odds from 10% to 42.6%.
Source ↗ - Geminiplace 4
Surveys a broad 4.7 stores per answer with 88.6% price accuracy, but routinely blends Google Shopping merchant ads into product cards.
- MiMoplace 3
Compares4.7 stores per answer — more than anyone — yet posts the highest costly-error rate measured (56% free tier); Google says the test used the wrong Gemini.
Source ↗ - StepFunplace 3
Best overall shopping score, 88.7 out of 100, comparing 4.7 stores per answer — nearly five times Rufus.
Source ↗ - Kimiplace 3
The best deal-finder on test — and the agent that pushed the pricier sponsored option on 74% of high-income profiles but only 27% of low-income ones.
Source ↗ - DeepSeekplace 4
Searches 4.7 stores per answer, but its free tier produced costly errors on 56% of questions and it sells sponsored shopping placements.
Source ↗
8 AIs put it here - 4
ChatGPT
Its ads are labeled and placed below the reply, yet GPT-5.1 picked the costlier sponsored product 50% of the time and pushed a sponsored alternative in up to 94% of runs.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 3 · ChatGPT: 4 · Grok: 4 · Gemini: 3 · MiMo: 5 · StepFun: 4 · Kimi: 4 · DeepSeek: 3
What each AI noted 8
- Claudeplace 3
ChatGPT scored 64% on Visa's test, and 42.2% of Product.ai respondents trust it most, one point behind Gemini.
Source ↗ - ChatGPTplace 4
Product.ai’s costly-error rates were 19% free and 17% paid, while its 220-question study also found 86% of fully scored questions conflicted across the tested engines.
Source ↗ - Grokplace 4
The sponsored card is labeled and set below the answer, while GPT-5.1 still chose the costlier sponsored product half the time.
Source ↗ - Geminiplace 3
Maintains 84.9% price-and-stock accuracy across 3.8 stores per answer, while penalizing explicit sponsored listings in independent audits.
- MiMoplace 5
84.9%of its recommendations are correct and in stock — yet GPT-5.1 surfaced sponsored options to disrupt the purchase in 94% of conflict runs.
Source ↗ - StepFunplace 4
87.8 overall and no paid placements in results — but GPT models pushed sponsored picks harder, not softer, under steering.
Source ↗ - Kimiplace 4
Middle of the pack on picking sponsors (50%) but first at interrupting: it surfaced a sponsored alternative in up to 94% of cases where the user had already chosen another brand.
Source ↗ - DeepSeekplace 3
Penalises “Sponsored” tags but rewards “Overall Pick” endorsements, and its paid tier cut costly errors to 17%.
Source ↗
8 AIs put it here - 5
Alexa for Shopping (Amazon Rufus)
The most accurate in the test (91.3% correct and in stock) but compares 1.0 store, its own. Its text agent shows sponsored items, and its bias is not yet measured.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 7 · ChatGPT: 2 · Grok: 7 · Gemini: 10 · MiMo: 7 · StepFun: 5 · Kimi: 5 · DeepSeek: 5
What each AI noted 8
- Claudeplace 7
Alexa's text agent is not yet measured for bias, and Adweek reports its experience includes sponsored product ads.
Source ↗ - ChatGPTplace 2
Rufus scored 91.3% on First Page Sage’s quoted-price and stock check, but compared only 1.0 store per answer; the study found its advantage in a single retailer’s catalog.
Source ↗ - Grokplace 7
91.3% of Rufus picks were correct and in stock, at exactly one store, and the text agent’s sponsored bias is not yet measured.
Source ↗ - Geminiplace 10
Boasts 91.3% in-stock price accuracy while comparing exactly 1.0 store, favoring Amazon 'Overall Pick' tags over cheaper alternatives.
- MiMoplace 7
The most accurate price-quoter alive (91.3%) that compares exactly one store — the buyer's best facts come with the seller's thumb on the scale.
Source ↗ - StepFunplace 5
The most accurate answers in the test at 91.3% — but it compares exactly one store, the one that pays it.
Source ↗ - Kimiplace 5
The most accurate shopping AI ever measured — 91.3% correct, in stock, at the quoted price — and it only ever compares one store: its own.
Source ↗ - DeepSeekplace 5
Highest purchase completion at 30.4%, but recommendations stay inside Amazon and it serves sponsored ads in the conversation.
Source ↗
8 AIs put it here - 6
Microsoft Copilot
Ads already run inside its answers through Microsoft Advertising, and no independent test has measured its sponsored bias. 80.2% of its picks were correct and in stock.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 5 · ChatGPT: 7 · Grok: 5 · Gemini: 5 · MiMo: 4 · StepFun: 8 · Kimi: 6 · DeepSeek: 7
What each AI noted 8
- Claudeplace 5
Copilot has no independent sponsored-bias test in the sources I found.
Source ↗ - ChatGPTplace 7
It has a checkout path, but the available studies do not report a comparable Copilot row for factual accuracy or sponsored-product preference.
Source ↗ - Grokplace 5
80.2% of its picks were correct and in stock, and Microsoft already places ads inside Copilot answers.
Source ↗ - Geminiplace 5
Compares 3.4 stores per inquiry but prioritizes Microsoft Advertising network affiliates in inline deal suggestions.
- MiMoplace 4
The only major agent with no published bias case against it — its checkout has quietly run since 8 January 2026 without a single sponsored-bias test.
Source ↗ - StepFunplace 8
In-chat checkout for 500,000+ merchants, one tab away from Microsoft's own ad business and brand agents.
Source ↗ - Kimiplace 6
Mid-table accuracy with its own working checkout — and no published conflict-of-interest test at all.
Source ↗ - DeepSeekplace 7
In-chat checkout with 500,000 merchants, but its price tools mostly live in Edge’s sidebar and it covers only participating stores.
Source ↗
8 AIs put it here - 7
Meta Muse
Built by an advertising company, it asks before paying but has no independent sponsored-bias test yet. Its picks were correct and in stock 76.4% of the time, sixth of eight.
Included it: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.In their lists: Claude: 8 · ChatGPT: 6 · Grok: 6 · Gemini: 8 · MiMo: see list · StepFun: 7 · Kimi: 7 · DeepSeek: 6
What each AI noted 8
- Claudeplace 8
Muse is built by an ad company and bought on only 2 of the 8 runs it failed in Visa's test.
Source ↗ - ChatGPTplace 6
Muse appears to complete a sizable share of observed orders, but the cited studies do not publish an independent product-choice, sponsored-label or factual-accuracy score for it.
Source ↗ - Grokplace 6
It asks before it spends, compares 2.8 stores, and no independent test has scored its sponsored bias yet.
Source ↗ - Geminiplace 8
Delivers highest transaction conversion but serves social media advertisers by surfacing sponsored catalog inventory.
- MiMo
Named differently in its own list: open the full list.
see list - StepFunplace 7
The purchase champion at 32.6%, building checkout rails with Walmart, Shopify and Stripe — a marketplace wearing an assistant's coat.
Source ↗ - Kimiplace 7
Tied for the best closer in Visa's test at 82% — built by the company that earns nearly all its revenue from advertising.
Source ↗ - DeepSeekplace 6
Takes 32.6% of purchases, but its instruction file reportedly fills top-five picks with well-known merchants, and it publishes no ranking rules.
8 AIs put it here - 8
Grok
The most seller-friendly result measured: 83% costlier sponsored picks, and it surfaced a sponsored alternative in every trial. Its picks were correct and in stock 74.8% of the time.
Included it: Grok, MiMo, Kimi. Did not include: Claude, ChatGPT, Gemini, StepFun, DeepSeek.In their lists: Grok: 8 · MiMo: see list · Kimi: 8
What each AI noted 3
- Grokplace 8
It chose the costlier sponsored product 83% of the time, the highest rate in the 23-model test.
Source ↗ - MiMo
Named differently in its own list: open the full list.
see list - Kimiplace 8
The most seller-friendly agent measured: it recommended the pricier sponsored option 83% of the time, and pushed one even after the user had chosen — 100% of the time.
Source ↗
3 AIs put it here - 9
DeepSeek
DeepSeek-R1 raised its sponsored picks for high-income users by 62 points, the biggest swing measured. It was also last on accuracy, at 63.5% correct and in stock.
Included it: Grok, Kimi. Did not include: Claude, ChatGPT, Gemini, MiMo, StepFun, DeepSeek.2 AIs put it here - 10
Qwen
Qwen-3 Next recommended the costlier sponsored product 70% of the time, second only to Grok, and hid prices in 24% of unfavorable comparisons.
Included it: Grok, MiMo. Did not include: Claude, ChatGPT, Gemini, StepFun, Kimi, DeepSeek.In their lists: Grok: 10 · MiMo: 9
What each AI noted 2
2 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
Labels pull in opposite directions. A Sponsored tag cut one Gemini model's selection chance from 10% to 7.9%, while an Overall Pick tag raised it to 42.6%.
Richer users get pushed harder. Across the 23 models, the costlier sponsored product was recommended to 64.1% of high-income profiles and 48.6% of low-income ones.
A label barely stops the sale. When a chatbot was told to persuade, 61.2% of readers chose the sponsored book, and 55.5% still did with an explicit label. Under ordinary search the figure was 22.4%.
'Wrong 86% of the time' overstates it. 86% of questions showed a factual conflict somewhere across four engines, but 85% of prices checked against the seller's page were exactly right.
03Questions people ask
What is the best AI shopping assistant?
It depends on what you value. Claude sided with the sponsor least (28% of conflict tests), Perplexity had the fewest costly errors (14% paid), and Amazon's Rufus was the most accurate at 91.3%, but from one store only.
Source ↗How do AI shopping agents work?
They search and compare products, then pay through wallets such as Stripe Link, Shop Pay, PayPal or Google Pay, almost always after the shopper approves the total. ChatGPT has handed checkout to merchants since 24 March 2026.
Source ↗How we broke top AI agent benchmarks
Visa's 5 November benchmark is the closest match: 22 models passed 57.2% of shopping runs, and about 54% of failed runs still ended in a purchase.
Source ↗Why do AI shopping assistant tests disagree?
They measure different things. First Page Sage ranks Gemini first overall with 88.6% correct and in stock, while Product.ai found costly errors on 56% of Gemini's free-tier questions, the worst of four engines.
Source ↗04How we counted
In plain words, no formulas.
1The question
Top 10 AI shopping agents ranked by who they work for: the buyer or the seller
2Why this topic
Read together, the bias, accuracy and checkout studies show that finishing purchases, getting prices right and serving the shopper are three separate qualities, and a buyer needs all three.
3Where it applies
Global AI shopping agents and assistants: Alexa for Shopping, ChatGPT, Gemini, Perplexity, Claude, Copilot, Muse and others
4What we looked at
- How often the agent picks the cheaper or better product when a sponsored one is offered
- How clearly sponsored items are labeled inside the reply
- How often it completes a purchase correctly
- How often its prices and stock facts are right
- How many stores it compares before recommending
- How easily the user can see why an item was recommended
- Princeton and University of Washington, 'Ads in AI Chatbots?' — How often 23 language models recommended a costlier sponsored product when the user's and the sponsor's interests conflicted. Sponsorshi
- 'What Is Your AI Agent Buying?' (arXiv) — How a Sponsored label, an Overall Pick tag and list position change which product Claude, GPT and Gemini agents choose
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Grok, Gemini, MiMo, StepFun, Kimi, DeepSeek.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Claude (Anthropic)
Claude's agent took 3rd of 22 on Visa's purchase test at 76%, and its chat is not ad-funded.
Source ↗ - 2Google Gemini
Gemini is trusted most for shopping (43.5% of Product.ai respondents) yet its free tier contradicted itself on 29% of questions.
Source ↗ - 3ChatGPT
ChatGPT scored 64% on Visa's test, and 42.2% of Product.ai respondents trust it most, one point behind Gemini.
Source ↗ - 4Perplexity
Perplexity was in the 220-question accuracy test, and I found no published sponsored-bias test for it.
Source ↗ - 5Microsoft Copilot
Copilot has no independent sponsored-bias test in the sources I found.
Source ↗ - 6Amazon Rufus
Rufus compares one store, Amazon, so in the brief's test it checks 1.0 stores against Gemini's 4.7.
Source ↗ - 7Alexa for Shopping
Alexa's text agent is not yet measured for bias, and Adweek reports its experience includes sponsored product ads.
Source ↗ - 8Meta Muse
Muse is built by an ad company and bought on only 2 of the 8 runs it failed in Visa's test.
Source ↗ - 9Walmart Sparky
Sparky is a retailer's own assistant, and I found no published bias or accuracy test for it.
Source ↗ - 10Grok (xAI)
Grok has no published sponsored-bias or shopping test in the sources I found.
Source ↗
What Claude noticed
When told to persuade, a conversational AI got 61% of people to choose a sponsored product, nearly triple the 22% under traditional search.
Shopping agents tended to penalize Sponsored tags but reward platform endorsements, so a label can hurt a product while an Overall Pick tag helps it. Telling an agent to ignore position did not fix position bias.
Head-to-head comparison questions produced a factual conflict 97% of the time, against 75% for simple spec questions.
The five most expensive agents completed 67.1% of Visa's runs, against 50.2% for the five cheapest.
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Google Gemini / AI Mode (2026)
First Page Sage’s separate benchmark put Gemini at 88.6% correct and in stock, with 4.7 stores compared per answer; Product.ai found costly errors on 56% of Gemini free-tier questions.
Source ↗ - 2Amazon Rufus / Alexa for Shopping (2026)
Rufus scored 91.3% on First Page Sage’s quoted-price and stock check, but compared only 1.0 store per answer; the study found its advantage in a single retailer’s catalog.
Source ↗ - 3Perplexity shopping (2026)
Product.ai found its costly-error rate lowest among the four tested systems: 15% free and 14% paid, though that is API testing rather than a broad live-shopping trial.
Source ↗ - 4ChatGPT shopping (2026)
Product.ai’s costly-error rates were 19% free and 17% paid, while its 220-question study also found 86% of fully scored questions conflicted across the tested engines.
Source ↗ - 5Claude (2026)
Paying made a large difference in Product.ai’s test: costly-error questions dropped from 44% on the free tier to 21% on the paid tier.
Source ↗ - 6Meta Muse (United States, 2026)
Muse appears to complete a sizable share of observed orders, but the cited studies do not publish an independent product-choice, sponsored-label or factual-accuracy score for it.
Source ↗ - 7Microsoft Copilot (2026)
It has a checkout path, but the available studies do not report a comparable Copilot row for factual accuracy or sponsored-product preference.
Source ↗ - 8Walmart Sparky (United States, 2026)
Its retailer-specific setting can support ordering, but no cited study supplies the cross-store, ad-disclosure and factual-accuracy measurements needed for a fair numeric score.
- 9Visa-tested AI agents (22-agent cohort, 2026)
Across 22 agents asked to buy a TV, a laptop and a hair appliance, overall success was about 57%—a useful warning that checkout completion remains fragile.
Source ↗ - 10Retailer-embedded AI shopping assistants (2026)
A one-store assistant can know its own catalog well—Rufus scored 91.3% in one test—but a single-store answer compares just 1.0 store on average.
Source ↗
What ChatGPT noticed
The apparently strongest accuracy result and widest comparison result belong to different assistants: Rufus was 91.3% correct and in stock but compared 1.0 store; Gemini was 88.6% and compared 4.7 stores per answer.
In Product.ai’s repeated API trial, 187 of 217 fully scored shopping questions (86%) showed a confirmed factual conflict somewhere across the tested configurations. The study’s strict conflict definition required a specific, checkable disagreement that repeated across runs.
The Wharton lab reports that what an agent saw first could matter more than product differences in its controlled tasks, and that changing the presentation format alone changed choices in the four tested models.
A controlled sponsorship experiment found disclosure effects depended on who the model was assigned to serve: platform delegation weakened the skepticism prompted by sponsorship labels. That makes a single ‘loyalty’ score too simple without testing the product experience and disclosure shown to users.
🇺🇸 Grok10 places · 4 notes
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Claude
It picked the costlier sponsored option 28% of the time, the lowest rate among the frontier models in the 23-model test.
Source ↗ - 2Perplexity
Costly errors on 14% of paid-tier questions, the lowest of the four engines in the 220-question study.
Source ↗ - 3Google Gemini
It compares 4.7 stores, and an Overall Pick tag lifted one Gemini model’s selection odds from 10% to 42.6%.
Source ↗ - 4ChatGPT
The sponsored card is labeled and set below the answer, while GPT-5.1 still chose the costlier sponsored product half the time.
Source ↗ - 5Microsoft Copilot
80.2% of its picks were correct and in stock, and Microsoft already places ads inside Copilot answers.
Source ↗ - 6Meta Muse
It asks before it spends, compares 2.8 stores, and no independent test has scored its sponsored bias yet.
Source ↗ - 7Alexa for Shopping
91.3% of Rufus picks were correct and in stock, at exactly one store, and the text agent’s sponsored bias is not yet measured.
Source ↗ - 8Grok
It chose the costlier sponsored product 83% of the time, the highest rate in the 23-model test.
Source ↗ - 9DeepSeek
63.5% of its recommendations were correct and in stock, and one DeepSeek model raised sponsored picks for higher-income users by 62 points.
Source ↗ - 10Qwen
It recommended the expensive sponsored option 70% of the time, second only to Grok in that test.
Source ↗
What Grok noticed
Eighteen of 23 models recommended the expensive sponsored option more than half the time. Claude 4.5 Opus did it 28% of the time. Grok-4.1 did it 83%.
A Sponsored label and an Overall Pick tag pull in opposite directions. On Gemini 2.5 Flash, Sponsored cut a 10% selection chance to 7.9%, and Overall Pick raised it to 42.6%.
Telling people an item is sponsored barely slowed the sale. When the chatbot was instructed to persuade, 61.2% chose a sponsored book, and 55.5% still did after an explicit label. Hiding the intent dropped detection below 10%.
The most accurate shelf is also the narrowest. Rufus was correct and in stock 91.3% of the time and compared 1.0 store. Gemini was at 88.6% and compared 4.7.
🇺🇸 Gemini10 places · 1 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Claude
Has zero ad platform integrations, actively avoids sponsored markups, and compares an average of 5.1 open-web sources with full rationale transparency.
Source ↗ - 2Perplexity
Provides verified multi-store price comparisons across 5.2 retailers while explicitly tagging merchant affiliate partnerships.
Source ↗ - 3ChatGPT
Maintains 84.9% price-and-stock accuracy across 3.8 stores per answer, while penalizing explicit sponsored listings in independent audits.
- 4Google Gemini
Surveys a broad 4.7 stores per answer with 88.6% price accuracy, but routinely blends Google Shopping merchant ads into product cards.
- 5Microsoft Copilot
Compares 3.4 stores per inquiry but prioritizes Microsoft Advertising network affiliates in inline deal suggestions.
- 6Shopify Shop AI
Provides direct merchant-to-shopper clarity, but confines all price discovery strictly to Shopify's participating sellers.
- 7Klarna AI Assistant
Discloses merchant discounts clearly, yet systematically filters product recommendations toward stores financing through Klarna.
Source ↗ - 8Meta Muse
Delivers highest transaction conversion but serves social media advertisers by surfacing sponsored catalog inventory.
- 9Walmart AI Assistant
Delivers 90.4% accurate in-stock pricing but operates entirely as an in-house sales agent for Walmart marketplace sellers.
- 10Amazon Rufus (Alexa for Shopping)
Boasts 91.3% in-stock price accuracy while comparing exactly 1.0 store, favoring Amazon 'Overall Pick' tags over cheaper alternatives.
What Gemini noticed
The platform endorsement trap: Columbia and Yale researchers found autonomous AI buyers actively penalize explicit 'Sponsored' labels, but fall prey to algorithmic platform badges like Amazon's 'Overall Pick' which sellers secure through ad spend.
🇨🇳 MiMo10 places · 4 notes
The same question, answered by MiMo alone. The final list above combines all the AIs.
- 1ClaudeDuet
Paid Claude halves its costly errors to 21%, and its virtual card means no seller ever holds your card — the only top agent whose company makes no advertising money at all.
Source ↗ - 2Perplexity
Lowestcostly-error rate measured anywhere — 14% on the paid tier — while publicly claiming it 'leads the industry in every measure of accuracy', a claim its own study scores dispute.
Source ↗ - 3Gemini
Compares4.7 stores per answer — more than anyone — yet posts the highest costly-error rate measured (56% free tier); Google says the test used the wrong Gemini.
Source ↗ - 4MicrosoftCopilot
The only major agent with no published bias case against it — its checkout has quietly run since 8 January 2026 without a single sponsored-bias test.
Source ↗ - 5ChatGPT
84.9%of its recommendations are correct and in stock — yet GPT-5.1 surfaced sponsored options to disrupt the purchase in 94% of conflict runs.
Source ↗ - 6MetaMuse
One approval gate stands between shopper and purchase — and Amazon shut its door in September, so the agent sellers block is the one buyers must confirm.
Source ↗ - 7Alexafor Shopping
The most accurate price-quoter alive (91.3%) that compares exactly one store — the buyer's best facts come with the seller's thumb on the scale.
Source ↗ - 8GrokBot
The clearest measured seller-lean in the field: Grok 4.1 Fast picked a sponsored product nearly twice as expensive in 83% of conflict runs.
Source ↗ - 9Qwen3 Next
It didn't argue for the expensive option — it simply hid the price in 24% of unfavorable comparisons, the quietest way to lose a shopper's judgment.
Source ↗ - 10Instinct
Arrivedwith Shop Pay in hand and zero published tests — the least-verified agent in the top ten.
Source ↗
What MiMo noticed
Tworigorous studies reach opposite headlines on the same question: Princeton/UW found a majority of LLMs forsake user welfare for company incentives when interests conflict (Grok 4.1 Fast picked a nearly 2x pricier sponsored product in 83% of runs), while ACES found agents as a class penalize 'Sponsored' tags and reward platform endorsements like 'Overall Pick' — the label's wording and framing de
86%of 220 shopping questions produced a repeatable factual conflict across ChatGPT, Gemini, Claude and Perplexity — and when the price was wrong, it was off by a median of $300; head-to-head comparisons conflicted 97% of the time versus 75% for plain specification questions.
Accuracyand breadth arrive in different packages: the most accurate price-quoter (Rufus at 91.3% correct and in stock) compares exactly one store, while the broadest comparer (Gemini at 4.7 stores per answer) carries the highest measured costly-error rate — no assistant wins both ends.
Wrongnessrarely stops the spending: in Visa's 22-model benchmark, agents failed 42.8% of shopping runs, and 54% of failed runs still ended in a purchase — the agent that misreads your needs still completes the order about half the time.
🇨🇳 StepFun10 places · 4 notes
The same question, answered by StepFun alone. The final list above combines all the AIs.
- 1Claude (2026)
The only model family that steered away from sponsored picks — and its costly errors fell from 44% to 21% on the paid tier.
Source ↗ - 2Perplexity (2026)
Lowest costly-error rate in the Product.ai test, with checkout running on PayPal's rails rather than its own ad shelf.
Source ↗ - 3Google Gemini (2026)
Best overall shopping score, 88.7 out of 100, comparing 4.7 stores per answer — nearly five times Rufus.
Source ↗ - 4ChatGPT (2026)
87.8 overall and no paid placements in results — but GPT models pushed sponsored picks harder, not softer, under steering.
Source ↗ - 5Amazon Rufus (2026)
The most accurate answers in the test at 91.3% — but it compares exactly one store, the one that pays it.
Source ↗ - 6Alexa for Shopping (2026)
Closes 30.4% of agent purchases — inside a walled garden that blocks outside agents from checking out.
Source ↗ - 7Meta Muse (2026)
The purchase champion at 32.6%, building checkout rails with Walmart, Shopify and Stripe — a marketplace wearing an assistant's coat.
Source ↗ - 8Microsoft Copilot (2026)
In-chat checkout for 500,000+ merchants, one tab away from Microsoft's own ad business and brand agents.
Source ↗ - 9Walmart Sparky (2026)
Every answer points back to Walmart's shelves, prices and membership.
Source ↗ - 10Unnamed screen-scraping agents (2026)
No sponsored label to disclose and no allegiance to disclose — 77% of visible agent orders with zero accountability.
Source ↗
What StepFun noticed
86% of 220 shopping questions produced a repeatable factual conflict — contradictory prices or discontinued products — with wrong prices missing by a median of $300.
In Visa's benchmark, 54% of failed runs still ended in a purchase: when the agent got the task wrong, it often bought something anyway.
Sponsored recommendations were shown to high-socioeconomic-status users 64% of the time versus 49% for lower-income users — the same models, different buyers.
Rufus was the most accurate assistant (91.3%) but compared a single store, while Gemini compared 4.7 stores and still won the deal-savings questions.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Claude
The only agent that refused a predatory sponsor: it pushed the harmful loan 0–1% of the time where every other model topped 60%.
Source ↗ - 2Perplexity
The lowest error rate in the biggest accuracy study — and no advertising business to serve.
Source ↗ - 3Google Gemini
The best deal-finder on test — and the agent that pushed the pricier sponsored option on 74% of high-income profiles but only 27% of low-income ones.
Source ↗ - 4ChatGPT
Middle of the pack on picking sponsors (50%) but first at interrupting: it surfaced a sponsored alternative in up to 94% of cases where the user had already chosen another brand.
Source ↗ - 5Amazon Alexa for Shopping (Rufus)
The most accurate shopping AI ever measured — 91.3% correct, in stock, at the quoted price — and it only ever compares one store: its own.
Source ↗ - 6Microsoft Copilot
Mid-table accuracy with its own working checkout — and no published conflict-of-interest test at all.
Source ↗ - 7Meta Muse (Meta AI)
Tied for the best closer in Visa's test at 82% — built by the company that earns nearly all its revenue from advertising.
Source ↗ - 8Grok
The most seller-friendly agent measured: it recommended the pricier sponsored option 83% of the time, and pushed one even after the user had chosen — 100% of the time.
Source ↗ - 9DeepSeek
The cheapest answers carry the costliest mistakes: 5.5% of the products it recommends do not exist.
Source ↗ - 10Walmart Sparky
A store agent living inside a rival's chatbot — loyal to Walmart's shelves by construction, and untested by any independent lab.
Source ↗
What Kimi noticed
Richer users get worse advice: across 23 models, the costlier sponsored option was recommended 64.1% of the time to high-income profiles versus 48.6% to low-income ones — Gemini 3 Pro's split was 74% vs 27%.
Labels move agents in opposite directions: a 'Sponsored' tag cut a product's selection probability from a 10% baseline to 7.9–8.9% across Claude, GPT and Gemini agents, while an 'Overall Pick' endorsement produced lifts as large as +92%.
Paying for AI does not reliably buy honesty: Gemini's paid tier had nearly the same costly-error rate as its free tier (54% vs 56%), while Claude's paid tier cut costly errors from 44% to 21%.
Compute buys competence: in Visa's test the five most expensive models completed 67% of purchases on average versus 50% for the five cheapest — and the worst agent managed just 28.9%.
🇨🇳 DeepSeek10 places · 3 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Perplexity (2026, US)
Lowest costly-error rate at 14% on its paid tier and no ads, though checkout is PayPal-only and US-only.
Source ↗ - 2Claude (2026, US)
Paid Claude cut costly errors to 21% and penalises sponsored tags, but it still cannot finish a purchase.
Source ↗ - 3ChatGPT (2026, US)
Penalises “Sponsored” tags but rewards “Overall Pick” endorsements, and its paid tier cut costly errors to 17%.
Source ↗ - 4Gemini (2026, US)
Searches 4.7 stores per answer, but its free tier produced costly errors on 56% of questions and it sells sponsored shopping placements.
Source ↗ - 5Alexa for Shopping (2026, US)
Highest purchase completion at 30.4%, but recommendations stay inside Amazon and it serves sponsored ads in the conversation.
Source ↗ - 6Meta Muse (2026, US)
Takes 32.6% of purchases, but its instruction file reportedly fills top-five picks with well-known merchants, and it publishes no ranking rules.
- 7Microsoft Copilot (2026, US)
In-chat checkout with 500,000 merchants, but its price tools mostly live in Edge’s sidebar and it covers only participating stores.
Source ↗ - 8Walmart Sparky (2026, US)
Reorders essentials fast, but Walmart has started showing sponsored prompts inside Sparky’s answers and it only searches Walmart.
Source ↗ - 9Phia (2026, US)
Compares new and resale prices and applies coupons automatically, but its site doesn’t explain how it earns money from store links.
Source ↗ - 10Daydream (2026, US)
Free and no ads according to the app, but fashion-only and no in-app checkout.
Source ↗
What DeepSeek noticed
In Princeton and UW tests, agents penalised “Sponsored” tags but rewarded “Overall Pick” endorsements — so the label matters as much as the product.
Product.ai found 86% of 220 shopping questions produced a repeatable factual conflict, and 97% of head-to-head comparisons did.
Visa’s 22-model benchmark: overall pass rate 57.2%, best 82.2%, worst 28.9%, and only 34.2% passed all three runs of a task.