In early August 2026, dozens of Amazon customers discovered that they could only see eight reviews on product pages — no sorting, no filtering, and no option to access the rest. There was no warning. Customer support explained that this was due to a violation of the terms of use regarding the collection and extraction of community content. In other words, the platform mistook real users for scrapers.
The story would seem like a mere curiosity if it weren't for its timing with another event: on August 4, 2026, the Ninth Circuit Court overturned the injunction that Amazon had imposed on the AI agent Perplexity Comet. The legal battle against automation had hit a snag — and almost simultaneously, the technical restrictions tightened significantly. Let's break down what exactly happened and why this changes the rules of the game for anyone collecting public data.
What Happened: Eight Reviews and a Five-Day Appeal
Mass restrictions were reported on August 3-4, 2026, by TechSpot, Android Authority, and Shopifreaks. The situation for those affected was the same:
- only 8 reviews are available for any product instead of the full list;
- sorting and filtering tools for reviews have disappeared;
- there is no notification about the restriction — users find out about it themselves;
- to lift the block, one must contact support and wait for a response for up to five business days.
One affected user, Fred Hall, encountered the restriction in mid-July while researching air purifiers. The chat operator cited a violation of the Conditions of Use regarding the extraction of user content; the appeal was approved a few days later without any explanation. A second case is even more telling: Sean Kaufman received the same limit of eight reviews across all products consecutively. He does not use VPNs, ad blockers, or AI browsers — his only distinguishing feature is that he spends a long time comparing competing products and carefully reading reviews. Amazon offered him $10 to his account, and access was restored after five days.
The company's official comment is as follows: "We understand that there may have been isolated cases where customers could not access reviews, and we have worked to ensure access." However, Amazon refused to disclose any substantial details: neither the number of affected accounts, nor the behavior that triggers the detector, nor the start date of the restrictions, nor the principles for reviewing appeals. Individual complaints, judging by discussions, have been reported since November 2025 — meaning the mechanism has been in place for a while, but only became noticeable in August.
Why Now: The Court Took Away Amazon's Convenient Tool
To understand the context, a timeline of the conflict between Amazon and Perplexity is needed.
- November 2024 — November 2025. Amazon claims to have sent Perplexity at least five warnings demanding that it cease agent purchases on behalf of users.
- August 2025. Amazon implements a technical barrier against Comet. According to Amazon, Perplexity released a workaround within 24 hours, disguising its agent as a regular Chrome session instead of honestly identifying itself.
- March 10, 2026. Judge Maxine Chesney in San Francisco issues a preliminary injunction: Comet cannot access password-protected sections of Amazon and make purchases on behalf of users. The key wording is that the agent acts "with the user's permission but without authorization from Amazon."
- August 4, 2026. The Ninth Circuit overturns the injunction. The court concluded that Amazon is unlikely to win under the Computer Fraud and Abuse Act: when a user delegates a task to an assistant, it is the user themselves who accesses Amazon's servers through the assistant, not Perplexity. The CFAA requires proof of intentional unauthorized access and damage exceeding $5,000 annually — elements that the court did not find.
Amazon responded that it "respectfully disagrees" with the ruling and reserves the right to seek a review. The case in the San Francisco court is not closed — only the preliminary injunction has been lifted. But the signal to the market has already been sent: the attempt to equate the actions of an agent acting on a person's direct instructions with hacking has not yet succeeded in court.
Perplexity, by the way, has insisted all along that the lawsuit protects not security, but advertising revenue: the agent purchases products bypassing the ad display that a human sees. This argument is not entirely far-fetched: Amazon's advertising business was valued at $56 billion as early as 2025, and all that revenue depends on a live buyer viewing the storefront themselves.
What Remains for the Platform When the Court Doesn't Help
Amazon has long been fighting on two fronts. In addition to legal battles, the company has consistently blocked access to AI crawlers via robots.txt: the list of blocked bots includes Meta, Google, Huawei, and Mistral, as well as earlier crawlers from Anthropic, Perplexity, and Google Project Mariner. In comparison, Walmart and eBay have not implemented such blocks in their robots.txt. The problem is that robots.txt is a request, not a barrier: it only works with those who voluntarily comply.
Thus, a logical step is to shift protection to a level where consent is not needed. This is precisely how the situation with reviews looks: instead of blocking specific bots by name, the platform begins to evaluate session behavior and limit output for those whose patterns resemble machine activity. Reviews are the most valuable and most copied asset of the marketplace, which is why the experiment began with them.
The practical downside of this approach is false positives. Behavioral classification relies on interaction dynamics: the rhythm and speed of movements, pauses, sequence of transitions, depth, and repeatability of views. A person who methodically compares ten models of air purifiers and opens all reviews for each one appears, based on these metrics, exactly like a parser. Kaufman's case is a textbook illustration: a clean IP, a regular browser, no automation, and still a ban.
The Main Conclusion: A Clean IP is No Longer a Pass
For those collecting public data, this story presents an unpleasant reality. Previously, the logic was simple: bad IP reputation — block, good reputation — access. Now Amazon has shown that decisions can be made despite perfect network signals. A user sitting with a home provider, from their own country, in a genuine Chrome browser, with a complete purchase history spanning years — and still faces restrictions because they read too many reviews in a row.
This changes the priorities in data collection infrastructure. A quality residential proxy is still necessary — without it, you won't even pass the first layer of filters and won't see the same output as a local buyer. But it is no longer a sufficient condition: if a robot operates on top of a good IP with a machine rhythm, the detector will catch you on the second layer. A detailed analysis of how this layer works can be found in our material on behavioral biometrics and proxies.
What to Do in Practice
- Physically separate "personal" and "work". If you monitor prices and reviews of competitors from the same browser and the same IP where you log into your buyer account, you risk both your data and your account. Use a separate profile, a different exit address, and a separate session.
- One pool — one task. Monitoring Buy Box, collecting reviews, and checking availability should not go through the same set of addresses: these tasks have different request patterns, and mixing them means creating a composite fingerprint that is more noticeable than any of the components.
- Limit depth per session. Eight reviews are not a random number: this is the threshold below which the platform considers interest normal. Collect reviews in batches through different sessions instead of pulling the entire feed in one go.
- Maintain a human rhythm. Random pauses, uneven transition speeds, and natural sequences (product page → reviews → back to results) are cheaper than rotating a thousand addresses in a rapid-fire request queue.
- Monitor degradation, not just errors. The most insidious aspect of Amazon's approach is that you don't get a 403. You receive an HTTP 200 with trimmed content. A parser that only checks the response code will diligently write eight reviews out of three thousand into the database for months without raising any alarms. Add completeness checks: number of elements, presence of pagination, existence of filter blocks.
- Build in a reserve. According to the State of Web Scraping 2026 study (Apify and The Web Scraping Club), 65.8% of specialists have started using more proxies than a year ago, and 58.3% have increased their budgets — even though prices per gigabyte have generally fallen. It's not the price that is rising, but the required volume and complexity. 43.1% of teams now maintain 2-3 providers for redundancy.
It is also worth keeping in mind the opposite trend: while some platforms mask filters, the industry is trying to legalize "good" bots — through cryptographic signatures of requests and explicit agent identification. We discussed how this works in our article on Web Bot Auth and signed agents. The problem is that a signed agent, by definition, gives itself over to the mercy of the platform: Amazon is more than happy to know that it is dealing with a bot — and will deny access.
Who Will Be Affected in the Coming Months
The quiet trimming of output instead of honest blocking is a tactic that can be easily replicated. It is likely that we will see it on other platforms: it is cheaper than CAPTCHA, does not spoil bounce metrics, and does not give the scraper a clear signal that they have been detected.
Those at greatest risk are scenarios where data completeness is important, not just access: review analysis for product analytics, reputation monitoring, training models on user content, competitive intelligence for sellers. If you work with a marketplace, it is useful to revisit basic practices for bypassing filters — for example, our analysis of Amazon's anti-bot protections when parsing prices — and add checks for response incompleteness.
This also affects ordinary users. A user mistaken for a bot loses exactly the function for which they came to the marketplace: the ability to read what others are saying about a product. There are no warnings, no explanations, no clear timelines for restoration — just a message to support and five business days of waiting.
Conclusion
August 2026 marked a turning point. The Ninth Circuit Court effectively stated that an agent acting on a person's behalf is a person, and cannot be subjected to computer abuse laws. Platforms understood this response and shifted their battle to where decisions are made without the courts — into the behavioral filters of their own frontends. The collateral damage in the form of live buyers limited to eight reviews seems to concern them moderately, judging by Amazon's comments.
The takeaway for practice is simple: the infrastructure for data collection is no longer a matter of "finding a good IP." A good IP is just an entry ticket, and from there, you are assessed based on your behavior and how much you request. Plan your budget, depth of requests, and completeness checks accordingly — and verify not only that data has arrived, but also that the amount is as expected.