Security software at 29 of 200 UK and US retailers, among them Home Depot, Best Buy, Macy's, Tesco, Argos and Next, refused an automated shopping agent sent by Blend Commerce on two separate attempts, even though every one of the refusing stores whose crawler rules could be read welcomes the search bots of ChatGPT, Perplexity and Claude. Blend, a conversion rate optimisation agency that works with Shopify merchants, published its AI Shopping Audit 2026 on October 6 and issued a statement on the findings on November 4, 2026, 23 days before Black Friday on November 27.

The audit scored each retailer out of 100 on whether an AI agent could reach the site, find products, read the product data and get a basket as far as the checkout. The average across all 200 was 65. Baby and nursery retailers led the 20 sectors with an average of 88; fashion came last with 49. Four retailers - Saks Fifth Avenue, Bergdorf Goodman, Anthropologie and JD Williams - scored zero because they refused every request Blend made, including requests for the home page and the robots.txt file.

In Short

A company that helps online shops sell more sent a piece of shopping software to 200 big and small stores in the UK and US to see whether it could get in, find products, read the product details and reach the checkout. One store in seven would not let it in at all, and the company blames security software, because every refusing store whose crawler rules it could read says AI search bots are welcome. The test used the company's own software rather than ChatGPT or Gemini, so it shows how stores treat automation they do not recognise, which is not quite the same as how they treat the big AI assistants.

How the audit was built

Blend set out four questions, according to the report: can an agent access the site, can it find the products, can it read and understand them, and can it get a basket to the checkout. The answers were combined into a single score.

The sample came from a researched pool of more than 900 retailers across 20 sectors. Every retailer in the pool sells a wide catalogue, either many brands or its own brand at scale; single-product brands, marketplaces, subscription boxes and trade-only sites were left out. In each sector and each country, Blend split the pool into thirds by estimated monthly organic visits and picked two large, two mid-sized and one emerging retailer. That produced 10 retailers per sector, five in each country, and a total of 80 large, 80 mid-sized and 40 emerging businesses. According to the report, "Platform was never part of the selection."

Testing ran between September 8 and 15, 2026. A companion survey of ecommerce leaders had run earlier, from July 15 to August 18. Nothing was bought. One item went into a basket and the journey stopped at the checkout page; on the endpoint test, Blend created one basket and one checkout and then cancelled both.

Four layers, 17 checks

The score rests on 17 checks grouped into four layers:

  • Reachable - a plain request to the home page from the store's own country and from the other country, plus a real browser loading a product page.
  • Discoverable - whether robots.txt allows the search crawlers used by ChatGPT, Perplexity and Claude, and whether the store publishes a product sitemap.
  • Understandable - product schema on the rendered page, with price, availability, GTIN, shipping details and returns policy.
  • Transactable - a live agent checkout endpoint, a basket taken as far as the checkout page, guest checkout, and express payment on the product page.

Each check carries a fixed number of points and the four layers add up to 100, according to the method. The report does not publish those weights. It says only that if a check could not be made, "it comes out of that store's total" - wording that does not settle whether an unmeasurable check scores zero or is removed from the maximum. That distinction matters most for the 25 refusing stores that still answered some requests.

How the agent presented itself is central to reading the results. "Plain requests gave the audit's name," according to the method, meaning they declared what they were. The browser was an unmodified build of Chromium, because, the report says, some content delivery networks spot a name added to a browser's user agent and refuse it. "We didn't impersonate a crawler or an AI company's agent at any point," the report states. The agent was, in other words, an honest but unfamiliar visitor - not ChatGPT, Gemini, Copilot, Perplexity or Meta's Muse.

Who refused, and why Blend blames security software

Twenty-nine of the 200 retailers - 14.5%, or one in seven - refused the agent. Blend defines a refusal narrowly: the store sent back an error, a challenge page or an empty page where the product page should have been, on both attempts, with each attempt coming from a residential connection in the store's own country. Only stores that refused twice were counted.

"Simple fact is, if an agent can't load your product page, it can't recommend what's on it," the report says.

The four zero scores sit at the hard end of the list. Blend says it re-tested Saks Fifth Avenue, Bergdorf Goodman, Anthropologie and JD Williams by hand and got the same result. "The score records the refusal. It's no judgement on the store," according to the report's frequently asked questions.

The robots.txt files said yes

The finding Blend says it did not expect concerns the crawler rules those same retailers publish. It could read robots.txt for 24 of the 29 refusing stores, and every one of the 24 allows the search crawlers from ChatGPT, Perplexity and Claude. "The refusals come from bot protection. That's the software that sits in front of the site and judges visitors by how they behave," according to the report.

Adam Pearce, co-founder and CEO of Blend Commerce, put the gap in organisational terms in the statement issued on November 4. "Most of these retailers have likely told the AI companies their agents are welcome. Then the security software on the front door refuses it. I'd bet most marketing teams don't even know it's happening," Pearce said.

"It's critical that companies test this for themselves ahead of the biggest shopping day of the year. Security and marketing teams need to get into a room to ensure overly strict security isn't costing thousands in sales," he added.

The report is less categorical than the statement on one point. The November 4 statement says that "not one of the 29 blocks AI in its robots.txt." The report's own count is that robots.txt was readable for 24 of the 29. Four of the remaining five are the zero-scoring stores, which refused even the robots.txt request; the report does not explain why the fifth file could not be read. For five of the 29, then, the audit holds no evidence either way.

The report does not dismiss the case for blocking. "Bot protection is there for good reasons. Nobody wants their website scraped without permission. Refusing can be the right call," it says, before asking: "But who made the call?" In Blend's experience, it continues, the decision usually sits with security or infrastructure teams "doing exactly what they were asked to do," without anyone consulting the person who owns revenue.

One in five large retailers

Refusals were concentrated among the biggest names. Of the 80 large retailers, 16 refused the browser (20%). Twelve of the 80 mid-sized retailers refused (15%), against one of the 40 emerging retailers (2.5%). The three tiers account for all 29.

A caption beneath that chart carries the audit's most important qualification: "A retailer may allow named AI agents while refusing automation it does not recognise." Bot protection products commonly keep lists of approved agents. Blend's test cannot show whether Home Depot or Tesco admits OpenAI's or Google's agents while turning away Blend's; it shows only that an unlisted automated visitor was refused. The report ends with an open invitation to the refused retailers to say "which agents you do let in."

Do big retailers score worse?

On average scores, the smallest retailers came top. Emerging retailers averaged 75, large retailers 63 and mid-sized retailers 61. That supports the idea that smaller retailers are better prepared than bigger ones only in part. It holds for the emerging tier, but mid-sized retailers scored lowest, two points below the large tier. The report itself notes that larger retailers were more likely to refuse the browser, "which lowers their average."

Blend offers a guess, labelled as such, for why. "We'd hazard a guess here that for the medium and big businesses, they are more likely to have the robust security teams who are stringent on bot-blocking," the report says.

The tiers carry their own caveat. Size is measured relative to each sector's pool, so "an emerging department store can be bigger than a large cycling retailer," according to the limitations section. The emerging tier is also the smallest group, at 40 retailers, two per sector.

Product data: prices yes, barcodes rarely

Blend could read product pages at 171 of the 200 retailers, reading up to three product pages per store in a real browser and counting a page only once it had confirmed it was a product page. Most published the basics in their structured data. Far fewer published the detail.

Of the 171:

  • 141 (82%) had product schema
  • 130 (76%) included price
  • 124 (73%) included availability
  • 38 (22%) included shipping details
  • 34 (20%) included a returns policy
  • 31 (18%) included a GTIN barcode number

Put the other way round, 41 of the 171 stores were missing price from their product data and 47 were missing availability.

The GTIN gap is the one Blend dwells on. A GTIN is the barcode number that identifies a product across retailers, and without it, according to the report, "an agent comparing prices has to guess whether two listings are the same thing." None of the beauty, fashion, luxury, pet or department store retailers whose pages Blend could read had one. Shipping and returns information, the report argues, is part of the buying decision; if it is missing from the product data, an agent has to hunt for it on a policy page or leave it out. "What helps a customer decide helps an agent decide," the report says.

Discovery was in better shape. According to the audit, 107 of the 200 retailers publish a product sitemap, leaving 93 without one that Blend could find.

Checkout: 38 endpoints, 37 of them on Shopify

The transactable layer produced the sharpest platform split in the study. Thirty-eight of the 200 retailers have a live agent checkout endpoint, which Blend defines as a published address where an agent can search the catalogue, build a basket and ask for a checkout without needing a browser. Thirty-seven of those 38 run on Shopify. The other is Albee Baby. The remaining 162 retailers had no endpoint.

Blend is open about its position. "Now, we're a Shopify agency. So yes, we would have a vested interest in pointing that out," the report says, adding: "And to be fair to everyone else, Shopify switches this on by default. Most of those merchants didn't lift a finger." For retailers on other platforms, the report points to the Universal Commerce Protocol as the standard to ask about.

That pattern - adoption arriving through a platform default rather than retailer choice - matches what PPC Land has reported elsewhere. Google launched the Universal Commerce Protocol on January 11, 2026, with Shopify among the co-developers. By May, Originality.ai had found only 26 UCP implementations, with the big names absent. In September, Google switched on AI Mode checkout for Shopify stores without asking, beginning on September 18 with an opt-out rather than an opt-in. Blend's audit window opened ten days before that rollout began.

Express payment on the product page was the one transactable feature most stores already had: 148 of the 171 readable stores offered it. Blend cites its own client tests in both directions - adding express payment lifted conversion by 35.8% for one client, and removing it lifted conversion by 13% for another.

The Shopify gap

The 37 Shopify stores in the audit averaged 91. The lowest scored 70, and four scored 100: Tanners Wines, Natural Baby Shower, Babythingz and Tiger Fitness. Retailers on every other platform averaged 59. The highest non-Shopify scores were Mattressman on 92 and Real Foods on 90.

Blend tested how much of that gap the free endpoint explains. With both endpoint checks removed from the scoring, Shopify stores still average 89, against 66 for everyone else, according to the report. The two sets of averages are consistent with the headline figure: by PPC Land's calculation, 37 stores at 91 and 163 at 59 produce an overall average of 64.9. Because the non-Shopify average rises when the checks are taken out, the remaining points must have been rescaled to 100. On that basis, and allowing for rounding, PPC Land estimates that the two endpoint checks carry roughly 9 to 12 of the 100 points; the method does not state their weight.

Shopify stores were not complete on product data either. Of the 37:

  • 36 (97%) had product schema
  • 34 (92%) included availability
  • 33 (89%) included price
  • 12 (32%) included a GTIN
  • 11 (30%) included shipping details
  • 10 (27%) included a returns policy

Thirty-three of the 37 were missing at least one of GTIN, shipping or returns, and 19 were missing all three. Blend traces this to the platform's defaults: according to the report, Shopify's built-in structured data covers name, brand, price and availability, but Shopify's own documented example does not include a GTIN, shipping details or a returns policy, so those fields appear only when a theme or an app adds them. All four 100-point stores had the full set.

Sector by sector

Each sector average describes ten retailers, five in each country, and the report warns that sector figures "describe the ten retailers audited, not the whole sector." With that limit in mind, the full ranking reads:

  • Baby and nursery: 88
  • Furniture and beds: 81
  • Jewellery and watches: 79
  • Cycling: 76
  • Health and wellness: 72
  • Musical instruments: 70
  • Video games: 70
  • Beauty: 69
  • Toys and hobbies: 69
  • Pets: 65
  • Homewares: 61
  • Alcohol: 59
  • Bathrooms, flooring and lighting: 59
  • Luxury: 58
  • Food and drink: 58
  • Garden and DIY: 57
  • Electronics: 55
  • Department stores: 51
  • Books: 50
  • Fashion: 49

Nine sectors sit above the all-retailer average of 65, one sits on it, and ten fall below. The report does not break the 29 refusals down by sector, so it is not possible to say how much of each low average comes from refused access rather than thin product data. Individual scores cut across sectors. In fashion, the lowest-scoring sector, 80s Casual Classics, an emerging UK retailer, scored 90. Toys R Us, a large US retailer, scored 94.

What 59 ecommerce leaders told Blend

Before the audit, Blend surveyed 59 ecommerce leaders between July 15 and August 18, 2026, on how ready they felt for agentic shopping and who owned the work. According to the results:

  • 63% say an AI shopping agent cannot currently complete their checkout
  • 61% have no 2026 budget for AI agent optimisation
  • 57% have nobody who owns AI readiness internally
  • 53% would not be surprised to learn they are already losing sales
  • 15% are not sure what an agent-ready checkout would involve

The survey and the audit cover different businesses, the report notes, and survey respondents are mostly brands with turnover below $10m. One figure does not reconcile with the stated sample. By PPC Land's calculation, 57% of 59 is 33.6 respondents; 33 respondents would be 55.9% and 34 would be 57.6%. The report does not say whether every respondent answered every question.

The demand side Blend points to

Blend's November 4 statement opens with a YouGov figure. According to the statement, planned AI use among UK shoppers "has almost doubled in a year - from 19% in 2025 to 34% in 2026." YouGov's report, published on October 8 under the title "The bots and the basket: AI's role in Black Friday and Cyber Monday 2026 in Great Britain," gives those figures for likely Black Friday and Cyber Monday shoppers rather than for all UK shoppers. The same YouGov page puts the share of Britons planning to shop in these sales at 23%.

The report adds seven studies from the past year, as summarised by Blend. According to that summary, Adobe Analytics found AI traffic to US retail sites rose 393% year on year in the first quarter of 2026, based on more than one trillion visits, and that in March 2026 such traffic converted 42% better than other traffic after converting 38% worse a year earlier; a later Adobe reading put AI-referred traffic up 138% year on year and 1,324% since October 2024. Salesforce found traffic from AI search doubled over Christmas 2025, with those shoppers converting nine times more often than shoppers from social media. Adyen's Retail Report 2026 found the share of UK shoppers using AI to shop rose from 12% to 28% in a year, based on 2,000 UK consumers surveyed by Censuswide in November and December 2025. Commerce and PayPal found 21% of UK online shoppers use AI tools to shop and 64% want to use agentic shopping tools, from 1,000 shoppers surveyed by Logica Research in March and April 2026. The Payments Association found 58% of 100 senior finance and risk decision-makers at UK online retailers believe they have already seen transactions made by AI agents. Bain & Company forecast in December 2025 that US agentic commerce could reach $300bn to $500bn by 2030, or 15% to 25% of US ecommerce.

Those figures measure different things. Most of the conversion data describes people who click through from an AI answer - AI referral traffic - rather than agents that browse and buy on a shopper's behalf. Blend's own definition separates three kinds of agent: those that answer product questions inside ChatGPT, Gemini, Copilot or Perplexity, those that open a site in a browser and click around like a customer, and newer ones that skip the browser and talk to a store through a checkout endpoint. Its audit tests the second and third. Evidence that the third kind converts at scale is thinner: Walmart said in March that purchases inside ChatGPT converted three times worse than click-outs.

For the US season as a whole, Adobe on September 28 put US online holiday spend at $275.1 billion, up 6.7%.

What the audit does not establish

Blend lists four limitations itself. Ten retailers per sector is a small group. UK and US stores were tested over different kinds of connections. The results are a snapshot from September 8 to 15, and stores change settings while platforms release updates. Size tiers are relative to each sector's pool.

The connection point contains an inconsistency the report does not resolve. The method section says every request came from a residential connection in the store's own country, and the refusals section says each attempt came from a residential connection in the store's own country. The limitations section says the UK used residential connections and the US used "a US proxy." Bot protection tools often weigh the type of network a visitor arrives from, so whether the 100 US stores saw residential or proxy traffic bears directly on the refusal count for the US half of the sample. The report does not break refusals down by country.

Several other gaps follow from the design:

  • Unrecognised automation, not named agents. The test establishes how bot protection treated an honest, unfamiliar automated browser. It does not establish how the same retailers treat agents from OpenAI, Google, Anthropic, Perplexity or Meta, which may be on allow lists.
  • Undisclosed weights. The points per check and per layer are not published, so a reader cannot rebuild any retailer's score. The report publishes scores only; the results of each of the 17 checks are offered to retailers in the audit on request.
  • No retailer responses. The report does not include comment from any of the 29 refusing retailers on whether the refusals were deliberate.
  • A commercial author. Blend sells conversion audits, and the report closes with an offer to run the same audit on other stores. It discloses its Shopify interest and tests whether the endpoint explains the Shopify lead, which narrows but does not remove that concern.

Why the refusals matter for marketers

The audit lands in a season in which AI shopping access has turned into a commercial fight as well as a technical one. Amazon blocked Meta's Muse agent from shopping on its site on September 20, citing among other things that Muse did not identify itself when browsing. That was a stated policy. Blend's argument is that many of the refusals it recorded are not policy at all but configuration nobody in marketing has looked at.

Identification is the crux. Arkose Labs found in October that 97.7% of legitimate AI agent sessions carry no verified ID. Cloudflare said on October 16 that it would give Verified bots trust scores for ignoring site rules, such as robots.txt instructions. Signed identity schemes such as Web Bot Auth are the route by which a named agent can prove who it is to a security vendor, and DataDome reported on October 30 that Meta's Muse agent had begun signing requests in a quiet test. Until signatures are common, most agents arrive looking much like Blend's did - automated and unverified.

The pressure on security teams runs the other way. DataDome counted AI requests to login pages rising from 11.9 million in January 2026 to 99.7 million in June, and found full bot protection on just 2.4% of popular websites. In a benchmark it ran itself, Resemble AI reported that six bot tools caught between 3.1% and 14.8% of disguised AI agents. The same benchmark put two of the tools above 99% for agents that did not hide their automation. Put beside Blend's audit, those vendor figures sketch an uncomfortable possibility: automation that declares itself is easy to stop, while automation built to look human gets through.

The checkout side is moving at the same time. Microsoft Clarity's data, published October 29, showed checkout appearing in 25% of AI agent sessions against 15% for humans. Shopify reported on November 3 that AI-referred orders rose 3.3 times as revenue gained 32% to $3.7bn, although it published no base for the AI figure.

Being visible in AI answers and being reachable by AI agents are also separate problems. Tesco, which appears on Blend's refused list, was ChatGPT's 25th most-cited UK domain in July, according to Semrush. A retailer can be recommended by an assistant and still turn that assistant's agent away at the door. For brands spending on visibility in AI search, Pearce's closing line in the November 4 statement makes the link explicit: "If you're spending money on being found in ChatGPT, Claude and wherever else, spend ten minutes before Black Friday checking an agent can actually get through your door."

Timeline

Summary

Who: Blend Commerce, a conversion rate optimisation agency working with Shopify merchants, and its co-founder and CEO Adam Pearce. The audit covers 200 UK and US retailers, including Home Depot, Best Buy, Macy's, Tesco, Argos and Next among the 29 that refused its agent, and Saks Fifth Avenue, Bergdorf Goodman, Anthropologie and JD Williams, which scored zero.

What: The AI Shopping Audit 2026 scored retailers out of 100 on 17 checks across four layers - reachable, discoverable, understandable and transactable. The average was 65. Twenty-nine retailers refused Blend's automated browser twice, yet all 24 refusers with readable robots.txt files allow AI search crawlers. Only 18% of readable stores publish a GTIN in product data, 38 of 200 have an agent checkout endpoint, and 37 of those run on Shopify, whose stores averaged 91 against 59 elsewhere.

When: The survey ran from July 15 to August 18, 2026, the audit from September 8 to 15, and the report was published on October 6. Blend issued a statement on the findings on November 4, ahead of Black Friday on November 27.

Where: 100 retailers in the UK and 100 in the US, across 20 sectors, tested from outside each store.

Why: Blend argues that bot protection set by security teams is turning away AI shopping agents that marketing teams want to reach, at a time when retailers are spending to appear in AI answers. The audit used an unrecognised automated browser rather than named agents from AI companies, so it shows how stores treat unfamiliar automation, not whether they admit ChatGPT, Gemini or Muse.