Amazon Ads has published the process it uses to keep Display, Video, and Audio campaign ads off unsafe web pages, and says fewer than 0.1% of impressions on the inventory it covers are served against content below its brand safety floor. The Support Center page, updated on November 16, 2026, states that the processes and the metric behind that figure have been submitted to the Media Rating Council for content-level brand safety accreditation, but "are not accredited as of the date of this document." It names one AI model provider: Anthropic, whose Claude foundation models Amazon evaluates "as part of model selection and ongoing measurement."
The brand safety page arrived three weeks after Amazon updated a second document, the general description of methodology for Display, Video, and Audio (DVA) campaigns, on October 26, 2026. That page sets out which DVA metrics carry accreditation, how each supply source is counted, and how fraudulent and non-human activity is filtered out. Read together, the two pages describe how Amazon's programmatic product, which replaced the Amazon DSP name at the end of September, measures and polices the ads it buys.
In Short
Amazon has explained, in two help pages, how it counts DVA ads and how it tries to keep them away from unsafe web pages, and it says fewer than 1 in 1,000 impressions in the covered inventory land on such pages. The brand safety process covers only English-language third-party websites, relies partly on AI models including Anthropic's Claude, and has not yet been accredited by the Media Rating Council. Every advertiser gets a fixed block on nine content categories that cannot be switched off, while apps, streaming TV, audio and non-English content sit outside the submission.
A submission, not an accreditation
The brand safety page is careful about its status with the Media Rating Council, the industry body that audits advertising measurement. Amazon calls the scope "Scope submitted for MRC accreditation but not accredited," and points readers back to the general methodology for metrics that are accredited. The submission, according to Amazon, "covers the processes Amazon Ads uses to keep advertising off content that does not comply with Amazon's Content Adjacency Policies (CAP)."
That scope is narrow. It covers one kind of inventory, on two supply sources, in one language.
| Brand safety submission | Status |
|---|---|
| English-language third-party web inventory, desktop and mobile web, through Amazon Publisher Services (APS) and third-party exchanges (3PX) | Submitted |
| Reporting of existing DVA metrics after the brand safety processes | Submitted |
| Third-party mobile application inventory | Excluded |
| Connected television, streaming television and audio inventory | Excluded |
| Content in languages other than English | Excluded |
| Programmatic Guaranteed deals on APS and 3PX inventory | Excluded |
| Brand suitability controls, including inventory tiers and content exclusion categories | Excluded |
Amazon's own properties - Amazon.com, IMDb, Twitch, Prime Video - do not appear in the submission at all. The process described is the one Amazon applies when DVA buys space on other companies' websites.
Within that scope, Amazon says brand safety is applied before money changes hands. "Display, Video, and Audio campaigns apply brand safety processes before bidding, so the metrics reported for that inventory inherently exclude content determined to fall below the brand safety floor as defined by CAP," the page states.
Nine categories, one floor
Amazon's Ads Trust team maintains the Content Adjacency Policies. According to the page, "Display, Video, and Audio campaigns block the following nine categories for every advertiser": Adult & Explicit Sexual Content; Arms & Ammunition; Hate Speech; Illegal Content; Terrorism; Violence, Death, & Injury; Accidents and Disasters; Online Gambling; and Alcohol, Drug Abuse, & Tobacco.
The list is not optional. "Every Display, Video, and Audio campaign within the submitted scope applies the CAP brand safety floor by default, which prevents ads from serving against content in any of the nine blocked categories. Advertisers cannot remove this default," the page reads.
Amazon says the categories "were developed in alignment with the Brand Safety Institute's (BSI) Brand Safety Floor and Suitability Framework," and publishes a mapping between the two.
| BSI category | CAP coverage of BSI floor, according to Amazon |
|---|---|
| Adult & Explicit Sexual Content | Adult & Explicit Sexual Content |
| Arms and Ammunition | Arms & Ammunition |
| Crime, Harmful Acts and Human Rights Violations | Violence, Death & Injury; Illegal Content; Hate Speech |
| Death, Murder, Willful Bodily Harm, Military Conflicts | Violence, Death & Injury |
| Online Piracy | Illegal Content |
| Hate Speech & Acts of Aggression | Hate Speech |
| Obscenity & Profanity | Hate Speech; Adult & Explicit Sexual Content |
| Illegal Drugs, Tobacco, E-Cigarettes, Vaping & Alcohol | Alcohol, Drug Abuse & Tobacco; Illegal Content |
| Spam & Harmful Content | Illegal Content |
| Terrorism | Terrorism |
| Debated Sensitive Social Issues | Harm-related aspects covered by Hate Speech and Violence Death & Injury; political issues covered through suitability (out of scope) |
| Misinformation | Harm-related aspects covered by Hate Speech, Violence Death & Injury, and Accidents & Disasters; highly debated social issues and politics covered through suitability (out of scope) |
Online Gambling does not appear anywhere in Amazon's mapping, so it is a category Amazon blocks beyond the BSI floor, and Accidents and Disasters appears only as partial cover for misinformation. Two BSI categories, meanwhile, are only partly covered. On misinformation, the page is blunt: "Amazon does not adjudicate whether a claim is true." Political content is handled through brand suitability tools, which are outside the submission.
Those suitability tools are what most advertisers will have touched. The page lists "inventory tiers, content exclusion categories, and advertiser-supplied blocklists for content above the CAP floor," which "let an advertiser exclude content that complies with Amazon's policies but does not suit their brand." Amazon DSP put its brand suitability tier settings into open beta in November 2025, with advertiser-level and ad-group-level controls available through the API, and added on/off content exclusion categories across 43 markets in December 2025. The new page places all of that above the floor, and outside what the MRC has been asked to review.
How the brand safety process works
According to Amazon, a page passes through eight stages between a property applying to carry ads and a bid being placed or withheld.
- Property eligibility. Amazon Ads assesses a publisher's web property before it can carry DVA advertising and keeps assessing it afterwards. The review covers brand safety and other quality criteria, "including whether the property carries unmoderated user-generated content." A property is either eligible or ineligible, and an ineligible finding "stays in effect until the publisher remediates and passes re-review through an appeal."
- Content capture. A crawler visits a page "after a Display, Video, and Audio campaign ad request is received with no current classification." Requests are counted over a rolling window, so the inventory carrying the most advertising is captured first. The crawler renders the page in a real browser engine, runs its scripts, and captures the first screen plus content below it.
- Language detection. The page's language is determined from the same render and stored with the classification. Advertisers can segment reporting by content language.
- URL screen. The page address is checked against "maintained keyword-and-context patterns that flag unsafe content."
- Two tuned models. Two models "built from published open models and tuned on samples of Amazon's advertising inventory" score the page, one on the address and one on the captured content, against each of the nine categories. Amazon also checks whether the page has changed and whether the capture worked.
- Two foundation models. "If either model identifies elevated risk," the page goes to two foundation models in sequence, each of which receives the page as an image. The first assigns policy categories, a risk level and the evidence it relied on. The second "reviews any finding from the first and removes findings it cannot confirm." According to Amazon, "A page is blocked only when a finding is confirmed."
- Vendor categories and industry lists. For every ad request, Amazon also consults third-party contextual vendors, and categories configured as unsafe are blocked. It keeps lists "derived from recognized industry sources for content such as child sexual abuse material and intellectual property crime."
- Pre-bid enforcement. When an ad request arrives, DVA checks property eligibility, domain-level blocks, the page's own determination and the vendor categories. "A bid is not placed on a request associated with content determined unsafe or for a property that is not eligible."
Content is reassessed on a recurring cycle, and frequently changing or higher-risk supply is checked more often. Blocks are sticky. "An unsafe determination does not expire, so a page found unsafe remains blocked," the page states, and only a publisher appeal changes it.
Amazon acknowledges the trade-off. "Because the check happens before an ad is served, it can reduce the inventory available to a campaign."
The crawler
The crawler is named. "The crawler identifies itself as AmazonAdBot rather than disguising itself," according to the page, which points publishers to adbot.amazon.com for the crawler's addresses and a way to verify it. That site gives the user agent string as "Mozilla/5.0 (compatible; AmazonAdBot/1.0; +https://adbot.amazon.com)", says the crawler honours standard robots.txt directives, and describes a reverse DNS check against the amazonadbot.com domain. It also warns publishers that blocking the bot will affect their access to Amazon Ads demand.
The methodology spells out what a block means on the measurement side. "Publishers can block the declared crawler through robots.txt, a challenge page, or an access gate. When Amazon Ads cannot capture content, it records the page as unassessed rather than safe," Amazon writes in its list of known limitations. Two further caveats follow: the crawler runs from infrastructure in the United States, so a site that varies content by location may show users elsewhere something different, and a property may serve the declared crawler a different version of its content than other visitors receive.
Claude and the models Amazon controls
The foundation model stage is where Anthropic appears. "The two foundation models are consumed as a service, and Amazon Ads controls the instructions given to them, their operating thresholds, and the versions in use," the page reads. In the section on business partners it adds: "Models such as the Claude foundational models used from Anthropic are evaluated as part of model selection and ongoing measurement."
The page does not say in so many words which of the two foundation models, or both, is Claude. It does make clear that the earlier, less resource-intensive tuned models are built from published open models, which Claude is not.
The relationship has a commercial side. Amazon invested $8bn in Anthropic convertible notes between the third quarter of 2023 and the fourth quarter of 2025, according to PPC Land's reading of Amazon's filings, and later converted part of that stake into preferred shares. The brand safety page does not mention the investment.
Amazon is frank about the weaknesses of generative classifiers. "Generative models can return different results for the same page on different evaluations," the page concedes, and Amazon says it reduces that risk by classifying in stages, confirming findings with a second model and measuring performance continuously. It also names a threat specific to the technology: "A page may also include text or imagery intended to influence a model's determination, and Amazon Ads maintains mitigations for this and tests them periodically." On bias, the page says foundation models "can reflect biases in their pre-training data" and that classification performance is measured by policy category to find uneven results.
The tuned models are trained on labels that a foundation model generates against "a category rubric written by Amazon," with human-reviewed sets validating those labels.
Content-level, by design
The approach matters in light of a rule the MRC tightened last year. In October 2025, the council barred property-level verification from being described as brand safety unless a vendor can also assess images, video and audio at the content level, with a grace period that ended on April 18, 2026. Amazon's pipeline sends each risky page to the foundation models as an image, "so it reads the page as a person looking at it would." Its own limits are content-level ones too: Amazon "does not independently assess content revealed through interaction or content in embedded video and audio players," and a page longer than the captured area "is assessed on the captured portion only."
Human review
Automated classification is backed by people, and the page sets out how their decisions are made final. "Reviewers apply the Content Adjacency Policies without seeing the automated system's determination, which keeps the review independent," it reads.
Two kinds of review run in parallel. Property review decides whether a site as a whole may carry ads. Page review takes individual pages that carried advertising and measures how the classification pipeline performed. Every determination comes with recorded evidence and a written rationale.
Escalation follows a fixed path. A finding that a property may not carry ads is checked by a second reviewer before it takes effect. Page-level findings are re-reviewed on a sampled basis. Where two reviewers disagree, a lead decides. Where a case "turns on the meaning of the policy rather than on the facts of the page, Ads Trust adjudicates, and that adjudication is final."
Changes to the system go through their own controls. Amazon says it records and controls changes to "policies, models, model instructions, operating thresholds, and reference sets," with peer-reviewed code review and a change management record naming approvers. Ads Trust signs off on changes that alter policy. Before release, a change's unsafe-content detection is compared with the version it replaces. A new model is chosen through a documented comparison of candidates, and Amazon "runs it alongside production without enforcing anything before it is permitted to block pages."
The 0.1% number
The headline figure comes from a metric Amazon calls the Brand Unsafe Impression Rate. "For inventory within the submitted scope, more than 99.9% of impressions are served against content that complies with Amazon's Content Adjacency Policies. Amazon Ads measures this through the Brand Unsafe Impression Rate, which is less than 0.1%," the page states.
The definition is given as a formula: "Brand Unsafe Impression Rate = impressions on unsafe content in sample / total impressions in sample."
How the sample is built determines what the number means. Amazon draws it from served impressions, so every page in it got past the blocking system. Most of the sample is impression-weighted; the rest is drawn at random from the lowest-volume inventory. Amazon explains why: "Impression-weighting alone would let a small number of high-volume pages determine the rate, while lower-volume inventory carries more brand safety risk, so the sample is built to represent both."
| Brand Unsafe Impression Rate measurement | According to Amazon |
|---|---|
| Disclosed rate | Less than 0.1% |
| Share of impressions on compliant content | More than 99.9% |
| Sample size, disclosed measurement | Approximately 1,000 pages a day |
| Measurement period | 30 days |
| Pages in disclosed sample | About 30,000 (PPC Land calculation from the two figures above) |
| Confidence interval | 99.7%, upper bound below 0.1% |
| Judgement | Trained reviewers on every sampled page; Ads Trust adjudicates |
| Daily internal measurement | One million pages a day, automated classification |
| Pages that cannot be assessed | Excluded from the rate, reported separately |
| Client notice trigger | Rate above 0.1% over a rolling 90-day period |
| Periods above 0.1% in past 12 months | None listed |
Each sampled page is crawled again for a fresh capture and assessed through "a measurement pipeline separate from the classification pipeline that blocks advertising." Trained reviewers check every page in the sample, and Ads Trust decides whether flagged content falls below the floor. That adjudication, Amazon says, is the basis for the rate.
A second, larger measurement runs every day on one million pages, using automated classification instead of reviewers. Human review tests each stage of that pipeline for false positives and misses. "The daily measurement supports internal monitoring. The rate disclosed above is the reviewer-judged measurement," the page states.
The nearest public comparison is YouTube. When YouTube Shorts gained its first MRC brand safety accreditation in June 2026, Google reported an Advertiser Safety Error Rate under 1%, based on 1,000 human-reviewed samples a day per format averaged over 60 days. The two figures describe different inventory, different formats and different floors, and are not directly comparable. On paper, though, Amazon's ceiling is a tenth of YouTube's, measured over half the window.
Exclusions shape the rate. "Pages that could not be assessed are excluded from the Brand Unsafe Impression Rate rather than counted as safe or unsafe, so the rate describes the inventory that could be assessed," according to the known limitations. Amazon says it measures and reports the size of that excluded portion separately. The page itself does not give that figure.
If the rate rises above the threshold, a public notice follows. "While Amazon Ads works to minimize the Brand Unsafe Impression Rate, clients will be notified in the event that it exceeds 0.1% over a rolling 90 day period, using a notice at the top of this article." The notice stays until Amazon has repeated the human-annotated study, and materially affected advertisers are told by email. For the past 12 months, the list of such periods reads "[None]".
What the October methodology accredits
The general methodology, updated on October 26, covers far more inventory than the brand safety submission. It is the document behind Amazon's MRC accreditation for DVA's core counting metrics, and it is explicit about where accredited numbers live: "Reporting of MRC accredited metrics is available within the Industry Standard custom report template in the Report Center."
| Measurement source | Creative | Device | Environment | Accredited metrics |
|---|---|---|---|---|
| Retail | Display, Video | PC, Phone, Tablet | Web, App | Gross begin to render impressions, Begin to render impressions, Gross click-throughs, Click-throughs, Playback initiation type (Video only), Viewability metrics |
| APS | Display, Video | PC, Phone, Tablet | Web, App | As Retail, plus Ad break position (Video only) |
| IMDb | Display | PC, Phone, Tablet | Web, App | Gross begin to render impressions, Begin to render impressions, Gross click-throughs, Click-throughs, Viewability metrics |
| 3PX | Display, Video | PC, Phone, Tablet | Web, App | As APS |
| Fire TV | Display | TV | App | Gross begin to render impressions, Begin to render impressions, Gross click-throughs, Click-throughs |
| Alexa | Display | Connected device | App | Gross begin to render impressions, Begin to render impressions, Gross click-throughs, Click-throughs, Viewability metrics |
The viewability metrics listed are measurable, viewable, non-viewable and non-measurable impressions (all rendered), plus measurable rate, viewable rate, viewable distribution, non-viewable distribution and non-measurable rate.
The exclusions are written as broadly as the inclusions. "MRC accreditation scope excludes any Measurement sources, Creative categories, Device types, Environment types, and Metrics not included in the table above. Accredited video metrics includes online video only and excludes streaming TV video. Any reporting other than the Industry Standard custom report template in the Report Center is excluded from MRC accreditation scope. Programmatic Guaranteed traffic from non-Amazon owned and operated publishers is also excluded from MRC accreditation scope."
The footprint has grown in stages. In May 2024, Amazon DSP added MRC accreditation for display begin-to-render impressions and click-throughs on third-party exchanges and Fire TV, again limited to the Industry Standard template. The October table lists six measurement sources, video on web and app, and a full set of viewability metrics. The page does not say when each element was accredited. Separately, Integral Ad Science secured MRC accreditation in November 2025 for its own third-party measurement of impressions, viewable impressions and invalid traffic on Amazon DSP.
Prime Video: submitted, not accredited
Prime Video sits in a separate table headed "Additional scope submitted for MRC accreditation but not accredited."
| Measurement source | Creative | Device | Environment | Submitted metrics |
|---|---|---|---|---|
| Prime Video Streaming TV Ads | Video | PC, Phone, Tablet, TV | Web, App | Gross begin-to-render impressions, Begin-to-render impressions, Gross clicks, Clicks, Playback initiation type, Ad break position |
| Prime Video Display Ads | Display | PC, Phone, Tablet, TV | App | Gross begin-to-render impressions, Begin-to-render impressions, Gross clicks, Clicks, Viewability metrics |
| Prime Video Live Events | Video | PC, Phone, Tablet, TV | Web, App | Gross begin-to-render impressions, Begin-to-render impressions, Playback initiation type, Ad break position |
| Prime Video Pause Ads | Display | TV | App | Gross begin-to-render impressions, Begin-to-render impressions, Gross clicks, Clicks |
Streaming TV video, in other words, is counted under a documented method but without the MRC's endorsement.
How Amazon counts an impression
The methodology describes a different counting trigger for almost every source.
| Supply | Format | Counting approach, according to Amazon |
|---|---|---|
| Retail, APS, IMDb, 3PX | Display, web and app | "count-on-begin-to-render"; pixel fires after elements of the creative have loaded |
| Fire TV | Display | "count-on-render"; a downstream service validates the ad rendered within the viewport |
| Alexa | Display | "count-on-render"; impression sent when the ad card rotates onto the screen |
| Prime Video | Display, including Pause Ads | "count-on-begin-to-render"; image loaded and minimum visibility criteria met |
| Retail, APS, 3PX | Online video | Counted once the video has begun playback on the client |
| Prime Video | Streaming TV video, including live | Counted once the playhead has progressed into the ad, from client-side signals |
For web and app display, the begin-to-render logic has known blind spots. Third-party creatives nested in extra iframe layers can prevent Amazon from detecting whether images loaded, in which case no begin-to-render impression is recorded. "This does not necessarily indicate that the ad did not render on the user's device," Amazon writes. Ad blockers, disabled JavaScript, disabled images and environments without the Page Visibility API also produce no counted impression. In mobile apps, some third-party bidders do not support the Open Measurement Software Development Kit, so Amazon "may be unable to determine whether playback began" for video.
Alexa gets unusual detail. Echo Show home screens rotate "up to 20 ads" alongside suggested content. Idle states begin "after 10 minutes without user engagement (presence detection or interaction), during which up to 8 ads may automatically rotate before idle activation," and no impressions are counted while the device is idle.
For living room devices, Amazon describes display-off detection. Apps running directly on a TV can read their own power state; set-top boxes and sticks rely on HDMI signals, and detection "is dependent on the TV panel's capabilities and the user's device configuration." Prime Video uses an "Are You Still Watching" prompt at content boundaries, and "Impressions are not served to inactive sessions."
Duplicates are removed at the bid level. "Every bid has a unique ID and if an incoming impression pixel has the same ID as a previously recorded impression, then this repeat impression is marked as a duplicate," the page states. Publishers are not paid for it, advertisers are not charged, and it does not appear in reports.
Clicks
Clicks follow one rule almost everywhere. DVA "enforces a one-click-per-impression methodology" across Retail, APS, IMDb, 3PX, Fire TV and Prime Video. Alexa is the exception: its ad cards can be shown several times, so it "uses a multi-click-per-impression methodology" with a refractory window between clicks, triggered by touch or a defined set of voice commands.
Alexa+ goes further. A click-through can also be triggered "by utterances outside the defined set when utterance intent matches the advertised action." Those clicks count in the accredited metric, yet, according to Amazon, "These undefined utterances have not been audited or accredited by MRC, and represent less than 1% of Alexa clicks."
Amazon also flags a source of mismatch with advertisers' own numbers. Because many redirects can sit between Amazon's click measurement and an advertiser's landing page, "This may lead to counting differences with downstream parties such as the advertiser."
Viewability, word for word
Amazon's viewability definition follows the industry standard. DVA "follows MRC and IAB defined viewability definitions, counting an ad as viewed when at least 50% of the ad shows on screen for 1 second or longer for display ads and 2 seconds or longer for video ads. For large size display ads which meet or exceed 242,500 pixels in size on PC devices, an ad is counted as viewed when at least 30% of the ad shows on screen for 1 second or longer."
| Ad type | Share of pixels on screen | Minimum time |
|---|---|---|
| Display | At least 50% | 1 second |
| Video | At least 50% | 2 seconds |
| Large display on PC (242,500 pixels or more) | At least 30% | 1 second |
The 242,500-pixel threshold is exactly the area of a 970 by 250 billboard, a PPC Land calculation.
Amazon says its scripts meet the IAB's polling intervals "of 100ms for display ads and 200ms for video ads," and that the count resets when an ad leaves focus. Some obstructions escape detection: picture-in-picture overlays, windows from other applications and notification bars may not report a change in focus, and "might result in a viewable impression being recorded when the obstruction occurs during initial visibility calculations." On Prime Video display, the client treats an ad as not visible when system dialogs or overlays cover it.
Invalid traffic: 5% to 10% blocked before the bid
The invalid traffic section carries the one hard number Amazon gives on filtering volume. "Amazon Ads' front-end techniques block approximately 5% to 10% of in-scope ad requests," according to the methodology. Pre-bid filtering uses deny lists and activity-based checks, supplemented by HUMAN Security, Jounce Media and DeepSee. Post-impression filtering runs in real time and through an offline retrospective process. Traffic caught before the bid is never bought, and traffic caught after the impression is not billed.
The accreditation covers general invalid traffic (GIVT) only. Amazon says it uses machine learning, network telemetry and client-side signals to detect automation tools, headless browsers and attempts to misrepresent traffic sources for select supply, but "Display, Video, and Audio campaigns is not accredited for Sophisticated Invalid Traffic (SIVT) filtration."
There is also a carve-out. GIVT is removed globally "except for Programmatic Guaranteed traffic from non-Amazon owned and operated publishers," an exclusion repeated in the census statement on IVT sampling.
Other figures from the section:
- External lists - the IAB/ABC spiders and bots list and TAG's Data Center IP List - are "monitored and updated on a monthly basis."
- Filtration runs across desktop, mobile web, mobile in-app and CTV "on a census basis. No subsets or sampling techniques are used."
- Amazon says "~99.9% of the traffic has necessary signals to make the IVT decision," so the decision rate "isn't reported discretely." That sentence still refers to the product as "ADSP", a trace of the old Amazon DSP name.
- Impression and completion beacons carry the cache headers "no-cache, max-age=0."
Advertisers who dispute IVT volumes can contact customer support. Where an issue is confirmed, Amazon processes refunds "according to materiality criteria," which the page does not define.
Reporting windows and retention
The methodology sets the timetable for when DVA numbers become final.
| Reporting rule | According to Amazon |
|---|---|
| Update frequency | Near real-time |
| Restatement from offline GIVT filtering | Within 72 hours, then "official and final" |
| Revisions for irregularity or miscalibration | Lookback period of 90 days |
| Raw impression and click logs | Retained for 30 days |
| Aggregated source data | Retained indefinitely |
| Reporting time zone | Time zone selected in the DVA account |
| Report delivery | Campaign Manager export; Report Center .csv to email or S3 bucket |
| Known reporting impacts, past 12 months | "No applicable notices logged in the past 12 months." |
Three metrics depend on publishers. Ad refresh rate is reported only when publishers pass auto-refresh information; otherwise, "Amazon Ads will not be able to detect or report on traffic generated as a result of auto-refresh." Paid traffic rate, which flags impressions from sourced-traffic arrangements, relies on UTM or referral information. Playback initiation type depends on a declared parameter. Amazon states that it does not use sourced traffic for Retail and IMDb.
Third-party providers
The methodology lists ten material providers.
| Provider | Role, according to Amazon |
|---|---|
| Trustworthy Accountability Group (TAG) | Data Center IP list to supplement invalid traffic identification |
| Interactive Advertising Bureau (IAB) | User agent spiders and bots list |
| Maxmind | Location databases to identify unknown or anonymous IP addresses |
| The Media Trust (TMT) | Identifies and reports ad creatives violating Amazon Ads policies |
| Integral Ad Science (IAS) | Pre-bid optimization integration on customer request |
| DoubleVerify (DV) | Optimization integration on customer request |
| HUMAN | Supplemental pre-bid blocking data for APS and 3PX |
| DeepSee | Supplemental pre-bid blocking data |
| Jounce Media | Supplemental pre-bid blocking data |
| RWS | Translation of DVA disclosures |
Pixalate, which Amazon DSP added for pre-bid invalid traffic filtering in January 2022, is not on the list. The brand safety page names DoubleVerify and IAS as providers of "supplementary content classification."
Where the pages meet what Amazon has said elsewhere
Two details in the October methodology sit awkwardly beside recent Amazon announcements.
The first concerns impression counting on Amazon's own store. On November 3, PPC Land reported a notice to advertisers in which Amazon described its current count for DVA display ads on the Amazon Store as count-on-download, and said that from January 1, 2027, impressions will count only once the ad appears on screen, about 63% fewer on average. The methodology describes Retail display differently. It says Amazon uses the "count-on-begin-to-render" approach, that "Amazon Ads' logic considers whether elements of the creative have loaded before firing the impression pixel," and that Retail impressions are only counted for billing and reporting when "the page serving the creative is completely loaded and available to the customer." Neither document reconciles the two descriptions, and the methodology page, dated October 26, does not mention the January change.
The second concerns where accredited numbers can be found. The methodology names one route for MRC-accredited metrics: the Industry Standard custom report template in the Report Center. That report is one of ten report types Amazon lists as not yet available in its new unified reporting system, and the old Display, Video, and Audio reports pages are scheduled to shut down on January 31, 2027, with saved reports and history deleted and editing frozen from December 17, 2026. The methodology does not mention unified reporting or the deadline. What happens to the only accredited view of DVA numbers if that report has no new home by the end of January?
Why it matters
For buyers, the brand safety page turns a set of claims into a documented process with a number attached. Brand safety claims have been contested across programmatic verification, and the MRC's response in October 2025 was to insist that they rest on content-level analysis. Amazon has submitted a pipeline that reads rendered pages as images, keeps humans as the final arbiters, and publishes a rate with a confidence interval and a public trigger for disclosure if it rises.
The limits matter as much. The submission covers English-language websites bought through APS and exchanges. It does not cover mobile apps, streaming TV, audio, Amazon's own properties, other languages, or the brand suitability controls advertisers actually configure. The rate excludes pages Amazon could not assess, and the page does not say how large that excluded share is. Nor has the MRC yet accredited any of it.
The use of foundation models is also new territory for an accreditation submission. Amazon has written down its own list of failure modes: inconsistent answers on the same page, pages engineered to sway a model, and bias inherited from training data. It names Claude, made by a company in which it holds a multibillion-dollar stake, as one of the models it evaluates. The page does not say how the MRC's review will treat a classifier whose answers can vary between runs.
For publishers, the practical message is in the crawler section. AmazonAdBot announces itself and respects robots.txt, but a page it cannot read is recorded as unassessed, and Amazon's crawler site warns that blocking it affects access to Amazon demand. In July 2026, Amazon Ads said it was already classifying user-generated content across more than 40 million pages a day into 20 sub-types. The new page explains what happens to a property that fails: removal from the eligible list until an appeal succeeds.
The measurement methodology, meanwhile, is a record of what Amazon will and will not stand behind. Retail, APS, IMDb, exchange, Fire TV and Alexa metrics carry accreditation; Prime Video's streaming and display metrics do not; SIVT filtering is not accredited; and the accredited figures live in a single report whose future home has not been announced.
What the pages do not say
The brand safety page does not give the size of the unassessed portion it excludes from the rate, the share of pages blocked at each stage, or the precise role Claude plays in the two foundation-model steps. It does not say whether DVA bids on a page while its first classification is pending. The methodology gives no date for each accreditation, no definition of the "materiality criteria" for IVT refunds, and no reconciliation with Amazon's January 1, 2027 impression change or the January 31, 2027 report shutdown.
Key facts
- Source documents: "Display, Video, and Audio campaign content-level brand safety description of methodology," Amazon Ads Support Center, updated November 16, 2026; "Display, Video, and Audio campaigns Description of Methodology," Amazon Ads Support Center, updated October 26, 2026.
- Accreditation status, brand safety: submitted to the Media Rating Council for content-level brand safety accreditation; "not accredited as of the date of this document."
- Brand safety scope: English-language third-party web inventory, desktop and mobile web, via APS and 3PX.
- Policy: Content Adjacency Policies (CAP), maintained by Amazon Ads' Ads Trust team; nine categories blocked for every advertiser; default cannot be removed.
- Framework: aligned with the Brand Safety Institute's Brand Safety Floor and Suitability Framework.
- Metric: Brand Unsafe Impression Rate, less than 0.1%; 99.7% confidence interval with upper bound below 0.1%; approximately 1,000 pages a day over 30 days, reviewer-judged.
- Crawler: AmazonAdBot, user agent "Mozilla/5.0 (compatible; AmazonAdBot/1.0; +https://adbot.amazon.com)", operating from US infrastructure.
- Models named: Claude foundation models from Anthropic, evaluated in model selection and ongoing measurement; tuned models built from published open models.
- Brand safety vendors: DoubleVerify and Integral Ad Science.
- Accredited measurement sources: Retail, APS, IMDb, 3PX, Fire TV, Alexa; Industry Standard custom report template in the Report Center only.
- Submitted, not accredited: Prime Video Streaming TV Ads, Prime Video Display Ads, Prime Video Live Events, Prime Video Pause Ads.
- Not accredited: Sophisticated Invalid Traffic (SIVT) filtration.
- Pre-bid IVT blocking: approximately 5% to 10% of in-scope ad requests.
- Restatement window: 72 hours; revisions lookback 90 days; raw logs retained 30 days.
- Source URLs: not shown on the captured pages; crawler information at adbot.amazon.com.
Timeline
- January 2022: Amazon DSP adds Pixalate pre-bid invalid traffic filtering across streaming TV, mobile and web supply.
- Q3 2023 to Q4 2025: Amazon invests $8bn in Anthropic convertible notes.
- May 2, 2024: Amazon DSP adds MRC accreditation for display begin-to-render impressions and click-throughs on 3PX and Fire TV, limited to the Industry Standard template.
- October 18, 2025: MRC bars property-level verification from brand safety claims without content-level capability.
- November 13, 2025: IAS secures MRC accreditation for third-party measurement on Amazon DSP.
- November 18, 2025: Amazon DSP brand suitability tier settings enter open beta.
- December 18, 2025: Amazon DSP adds content exclusion categories in 43 markets.
- April 18, 2026: MRC grace period for property-level vendors ends.
- June 3, 2026: YouTube Shorts gains MRC brand safety accreditation, with an Advertiser Safety Error Rate under 1%.
- September 29, 2026: Amazon replaces the Amazon DSP name with DVA.
- October 26, 2026: Amazon updates the DVA campaigns Description of Methodology.
- November 3, 2026: PPC Land reports Amazon's notice moving Amazon Store DVA display impressions to an on-screen count from January 1, 2027.
- November 12, 2026: PPC Land reports the January 31, 2027 shutdown of legacy report pages, with the Industry standard report not yet in unified reporting.
- November 16, 2026: Amazon updates the DVA content-level brand safety description of methodology, reporting a Brand Unsafe Impression Rate below 0.1%.
- December 17, 2026: Legacy report pages scheduled to become read-only.
- January 1, 2027: On-screen impression counting scheduled for DVA display ads on the Amazon Store.
- January 31, 2027: Legacy Display, Video, and Audio reports pages scheduled to shut down.
Related PPC Land coverage
- Amazon drops off-screen ads from impression counts, expects 63% fewer - Amazon's notice on moving Amazon Store DVA display impressions to an on-screen count from January 1, 2027.
- Amazon advertisers will lose unmigrated reports on January 31, 2027 - The ten report types, including the Industry standard report, not yet available in unified reporting.
- Amazon calls its ad platform an agent, but the agent is still closed beta - The September 2026 renaming of Amazon DSP to DVA.
- Amazon DSP bolsters measurement transparency with additional MRC accreditation - The 2024 addition of accredited display begin-to-render impressions and click-throughs on exchanges and Fire TV.
- IAS earns MRC accreditation for third-party Amazon DSP measurement - IAS accreditation for impressions, viewable impressions and invalid traffic on Amazon DSP.
- Amazon DSP launches brand suitability settings in open beta - Advertiser and ad-group level suitability tiers through the API.
- Amazon DSP adds content exclusion categories for brand suitability control - On/off exclusion categories for Twitch and third-party web and mobile inventory in 43 markets.
- Scope3 launches AI brand safety integration with Amazon DSP - An earlier AI-based brand safety option offered to Amazon DSP buyers.
- MRC restricts property-level ad verification from brand safety claims - The October 2025 policy requiring content-level capability for brand safety claims.
- YouTube Shorts gets its first MRC brand safety accreditation - a short-form first - Google's accredited Advertiser Safety Error Rate and its sampling method.
- Amazon DSP introduces Pixalate pre-bid invalid traffic filtering - The 2022 pre-bid IVT integration across streaming TV, mobile and web.
- Amazon, Google and Microsoft hold $269bn of Anthropic compute contracts - Amazon's investment in Anthropic and its compute commitments.
- IAS gains ChatGPT Ads safety pilot without access to user conversations - Verification vendors testing brand suitability in AI chat environments.
Summary
Who: Amazon Ads, through its Display, Video, and Audio (DVA) campaigns product and its Ads Trust team, with the Media Rating Council as the accrediting body, Anthropic as the named foundation model provider, and DoubleVerify and Integral Ad Science as classification vendors.
What: Two Support Center methodology pages. The brand safety page describes a crawler, model and human-review pipeline that blocks nine content categories before bidding and reports a Brand Unsafe Impression Rate below 0.1%, submitted for MRC accreditation but not yet accredited. The general methodology sets out which DVA metrics are accredited, how impressions, clicks and viewability are counted, and how invalid traffic is filtered.
When: The general methodology was updated on October 26, 2026, and the brand safety methodology on November 16, 2026.
Where: The brand safety process covers English-language third-party websites, desktop and mobile web, bought through Amazon Publisher Services and third-party exchanges. The measurement accreditation covers Retail, APS, IMDb, 3PX, Fire TV and Alexa, with Prime Video submitted but not accredited.
Why: Advertisers and the MRC have pressed for content-level evidence behind brand safety claims, and the documents give DVA buyers definitions, thresholds and exclusions they can check against their own reports, at a time when Amazon is changing its impression counting and its reporting tools in January 2027.
Discussion