Half of web requests are bots as agency CPMs climb 20%

Two clients of the same agency looked at the same traffic report this quarter and reached opposite conclusions. One saw a distribution problem worth solving. The other saw a bill worth disputing. Neither was wrong, and that is the difficulty.

Digiday published research on September 8, 2026 documenting a split inside the buy side over what to do about the machines now arriving at brand websites. Cloudflare data cited in the piece puts bots and agents past 50 percent of all web requests. GoFish reported clients seeing bot traffic rise 80 percent year over year. A direct-to-consumer founder in the same article described artificial-intelligence assistants sending traffic that converts several times better than his site average, and wants more of it.

The week produced three other stories that share the shape of that argument. Google published documentation for a feature that reads promotional offers off a merchant's website and inserts them into advertisements without the advertiser writing them. A search relations representative explained that Google's systems can lose faith in a site that generates pages by combination. TikTok Shop banned the printed slip of paper that asks a buyer for five stars. And an Australian sports broadcaster published a 36 percent reach figure it will not be able to verify independently until 2027.

The common thread is not artificial intelligence. It is the widening gap between a signal and the thing the signal was supposed to represent, and the growing number of parties with an incentive to manufacture the signal rather than the thing.

Half the requests, and no agreement on what to do about them

The dilemma Digiday reported on September 8 does not follow the usual line between advertisers and platforms. It runs between advertisers.

On one side sit brands that treat agentic traffic as a distribution channel. Chad Keller, co-founder of the mattress company Mellow Sleep, told Digiday that artificial-intelligence assistants are already sending traffic that converts several times better than his site average. The British retailer John Lewis has watched agentic searches grow from 0.3 percent of visits to 2.5 percent over a year and is actively seeking more of them. Those are small percentages attached to a large multiple, which is exactly the profile that makes a channel worth investing in early.

On the other side sit the agencies that buy media against the resulting data. David Dweck, president of GoFish, described clients seeing bot traffic increase 80 percent year over year, and reported that the adjustments made to retargeting in response pushed cost per thousand impressions up by an average of 20 percent. One client saw cost per acquisition spike temporarily in the fourth quarter and into the current year. Practitioners quoted alongside him included Mallory Beck, vice president of client services at Lighthouse Creative, Megan Herling, executive director of media at LERMA/, Nola Ladd, brand media supervisor at Collective Measures, Colin Maduzia, senior vice president of experience and product strategy at RPA, Justin Scarborough, head of programmatic at Crossmedia, and Tim Lathrop, vice president of platform digital at Mediassociates.

The mechanism connecting the two positions is the retargeting pool. A retargeting audience is built by recording which visitors reached which pages, then buying advertising against that recorded list. When a substantial and growing share of the visitors doing the reaching are not people, the list stops describing a population of potential buyers and starts describing a population of requests. The advertiser then pays a real price to reach an audience whose composition has quietly changed.

Two things follow from that, and both showed up in the reporting. Prices rise, because the pool has to be filtered, narrowed or replaced, and each of those operations costs money in the auction. And budget moves, because the simplest response to a data quality problem on the open web is to spend somewhere the data quality problem is somebody else's. Digiday reported clients shifting spend toward social platforms and retail media networks, both of which are closed environments where the identity of the visitor is established by a login rather than inferred from a request.

That shift is worth stating plainly, because it is a competitive consequence rather than a technical one. Open-web publishers do not control how many bots arrive at their pages. They do carry the cost when the resulting audience data is treated as unreliable, and the beneficiaries are the environments that never depended on that data in the first place.

The volume is not new, and neither is the measurement problem underneath it. Cloudflare has documented artificial-intelligence crawlers hitting individual sites 50,000 times per referred visitor, and separately reported that automated bots have overtaken humans across its network. Kinsta measured artificial-intelligence bots hitting WordPress cart pages 3.75 million times in a single day, which is a specific and instructive number because cart pages are precisely the pages a commerce retargeting pool is built from. Decodo's analysis found the United States sending 53.5 percent of global bot traffic. Crawler share of overall web traffic reached 4.2 percent as the internet grew 19 percent in 2025, a figure that measures crawling rather than the wider category of automated requests.

The gap between the scale of the problem and the response to it is documented too. Lunio found only 5.3 percent of marketers using invalid traffic tools while 75.6 percent said they lose advertising budget to bots. The same firm measured 72 percent more invalid traffic in retail search campaigns running on AI Max, which places some of the automation inside the buying stack rather than outside it. Invalid traffic as a category has an established definition and an established verification industry, and DoubleVerify has reported ad fraud falling 41 percent in North America and 45 percent in EMEA. Fraud detection improving and bot volume rising are not contradictory findings. They describe different things. Detection catches the traffic that is trying to steal money. The traffic Dweck's clients are absorbing is mostly not trying to steal anything: it is crawlers, agents and assistants doing legitimate work, which is what makes it hard to filter and impossible to invoice back to anyone.

There is also a genuine measurement question buried in the optimistic half of the argument. Traffic that converts several times better than site average is a description of quality, but it is also a description of intent that has already been formed elsewhere. An assistant that sends a visitor to a mattress site has typically done the comparison, the filtering and the shortlisting before the click. Invoca has measured calls originating from ChatGPT converting to leads at 49 percent, ahead of every other channel, which is the same pattern in a different medium. High conversion on referred traffic is compatible with the referring layer capturing the part of the funnel that advertising used to be paid for. Whether that is a channel worth investing in or a channel that quietly reprices everything upstream of it is not a question either side of the Digiday piece answered, and it is not obviously answerable from a conversion rate alone.

Shopping ads got fewer impressions and more clicks

The second story of the week runs in the opposite direction from the first: a case where the numbers improved and the reason is unflattering.

Mike Ryan of Smarter Ecommerce published an analysis on LinkedIn, covered by Search Engine Roundtable on September 7, drawn from a dataset of 175 billion impressions spanning mid-2025 to mid-2026. Median impressions for Google Shopping advertisements fell from roughly 1.85 million to 1.4 million. Median click-through rate rose from 1.20 percent to nearly 1.55 percent.

Ryan called it a reverse crocodile effect. The original crocodile effect, sometimes called the great decoupling, describes what happened to organic publishers after AI Overviews launched: impressions climbing while clicks fell, the two lines on a chart opening like a jaw. Shopping advertisements produced the mirror image. Fewer impressions, better conversion of the impressions that remained.

His proposed explanation is the part worth sitting with. Ryan theorised that Google selectively serves AI Overviews on lower-probability queries in order to preserve revenue, which would mean the impressions removed from the Shopping pool were disproportionately the ones least likely to have produced a click anyway. Click-through rate then rises from both directions at once: a smaller denominator, and a denominator with the weakest members taken out of it.

If that reading holds, a rising click-through rate in a Shopping account over the past year is not evidence that anything improved. It is evidence that a selection decision was made upstream, by a party that also benefits from the resulting number looking better. Nobody at the advertiser changed a bid, a feed or a creative to produce it.

The organic side of the comparison has been measured repeatedly and by different methods, which is why the contrast is usable at all. A randomised study found AI Overviews cutting publisher clicks 39.8 percent. Separate work has put the organic click reduction at 34.5 percent, and other measurement has found organic click-through rate down 61 percent and paid traffic down 68 percent in the affected result sets. The spread between those figures is wide enough that no single number should be quoted as settled. The direction is consistent across all of them.

Two Google Ads changes surfaced in the same 48 hours, and both move a decision from the advertiser to the system.

Hana Kobzová of PPC News Feed spotted a new help document on September 7 describing automated promotions as an account-level automated asset. The feature reads promotional offers off an advertiser's own website, something along the lines of 20 percent off black sweaters, and displays them inside Search and Performance Max advertisements. Google's language is that it highlights valid, high-quality deals and offers directly within ads to drive both online and offline conversion discovery. Eligibility requires an active Search or Performance Max campaign with location assets attached, no existing manual promotions on the campaign, and policy compliance. The guidance for advertisers who want control is to keep website promotional details current, or to upload specific promotions manually. The document Kobzová surfaced does not describe an explicit opt-out.

That is a meaningful design choice. An advertisement carrying a specific discount claim is a legal representation as well as a marketing one, and the source of truth for that claim is now a page on the advertiser's site being read by a parser. Expired offers, regional pricing, conditional discounts and promotional pages left up after a sale ended are all ordinary features of retail websites and all potential inputs. Google has added promotion capabilities repeatedly over the past year, including audience targeting for Shopping promotions in Merchant Center. The distinguishing feature of this one is that nobody has to write anything.

The second change moves the other way. PPC Land reported on September 6 that Lookalike segments reached Video brand campaigns through documentation published September 1, carrying three explicit reach tiers: narrow at 2.5 percent of users in the target location, balanced at 5 percent as the default, and broad at 10 percent. Seed lists require more than 100 active matched people according to the documentation, while a separate Google page recommends 1,000 for quality, a discrepancy the article flags rather than resolves. Acceptable seeds include Customer Match lists, website and application engagement data, and YouTube content interactions. Segments refresh every one to two days, can only be created during campaign setup through the ad group's audience section, and cannot be edited afterwards in Audience Manager.

The same product name now behaves differently depending on where it is used. Video campaigns enforce those percentages as hard reach caps. Demand Gen, which lost its hard targeting control in March 2026, treats the same percentages as directional signals fed into the system rather than limits. An advertiser reading a single Lookalike help page has no way to know which behaviour applies without knowing the campaign type first.

There is a connection back to the opening story that goes unremarked in the documentation. Website and application engagement data is one of the three permitted seed sources. A seed list assembled from site engagement inherits whatever the site's traffic actually consists of, and the modelled expansion then multiplies the composition of the seed by a factor of somewhere between 25 and 100.

Programmatic pages, and the trust that takes time to rebuild

John Mueller of Google addressed programmatic SEO on Bluesky on September 7, and Search Engine Roundtable covered the exchange the same day.

The technique in question generates large numbers of pages by iterating through combinations: a domain name crossed with a technology crossed with an attribute, each combination becoming its own URL. It is cheap, it scales without editorial cost, and it has been a standard tactic in software marketing for years.

Mueller's assessment was that it often leads to a site that is either spam, borderline spam, or low quality. Individual pages may carry some value, he allowed. The difficulty is the aggregate. It is easy to spin something up with many pages, he said, and hard to provide real value to users.

The consequential part was not the diagnosis but the remedy, or the absence of one. Once the threshold is crossed, in Mueller's phrasing, Google's systems have possibly lost faith in the site providing good value to users. Resolving that tends to take time and significant effort to show the value. There is no reconsideration request for a site that has simply stopped being trusted, because no manual action was applied. A manual action has a defined appeal path. A quiet reduction in systemic confidence has none, which makes recovery indistinguishable from the ordinary work of building a site up in the first place.

This is the same enforcement architecture that produced the results in Google's third spam update of 2026, which rolled out to every language and region, and it sits alongside independent findings that commodity content is losing its place in the index outright. What Mueller described is not a penalty. It is a downgrade in a probability estimate, applied to a whole domain, with no notification and no defined path back.

Apple made an adjacent clarification over the same weekend. Its Applebot support documentation, timestamped September 7, added a line stating that site rules for Applebot-Extended are not considered in ranking for Search. Search Engine Roundtable reported the change on September 7. Applebot-Extended is the directive that governs whether a site's content feeds Apple Intelligence and Siri features rather than Apple's search index, and Apple has now stated explicitly that using it carries no ranking consequence in either direction. The wording mirrors Google's position on Googlebot-Extended almost exactly. Apple rewrote the Applebot rules to feed Siri's artificial-intelligence features earlier in this cycle and has since added 4,656 IP addresses to the crawler in a single update, so the operational footprint is expanding while the stated ranking consequence is being narrowed to zero.

TikTok Shop bans the slip of paper that asks for five stars

TikTok Shop published packing standards in its US Academy on September 3, 2026, and PPC Land covered the ten prohibited practices on September 6.

Two of the ten address the physical mechanics of manufacturing a rating. Review-request messages inside packages are prohibited, and so are materials designed to incentivise positive reviews. The rest of the list is closer to conventional logistics policy: dropshipping directly from other retailers, including unpurchased products, unauthorised marketing materials such as pamphlets, price tags and stickers, loose-fill materials where local requirements prohibit them, bundle strapping that impedes handling or scanning, large staples or nylon tape creating safety hazards, non-original pesticide packaging, and non-corrugated outer packaging for hazardous materials.

The packaging requirements themselves are specified by category and by count of controls. Standard items divide into six categories with descending protection requirements: spillable items require seven controls, perishable items five, fragile or breakable items four, potentially breakable items four, high-value items three, and non-breakable items one. Bulky and irregular items divide into three categories. Fragile, bulky or heavy items require all six controls. Non-fragile bulky or irregular items require cushion, filler and a wooden crate. Liquid items in manufacturer's packaging require four.

One detail in the published policy is worth noting for what it reveals about the drafting process. The bulky category definition contains an unresolved editorial note reading that thresholds need US validation before publishing, attached to the 10 kilogram and 100 centimetre cutoffs. That note appeared in the live policy. Sellers reading the definition are therefore reading a threshold the platform has not confirmed, in a document that governs their compliance.

Enforcement runs through the scoring systems TikTok Shop has been assembling for most of a year. Shop Performance Score affects affiliate access and settlement speed. Account Health Rating blocks new listings and campaigns when a seller falls to 150 points. Damaged packaging counts as seller fault in the return rate calculation, which means a packing failure feeds a metric that gates advertising eligibility. The paid and incentivised review prohibitions were established in June 2026 alongside a 60-day negative review rate window, and the policy sits within a broader tightening that includes a ban on minors selling in creator videos and livestreams and shipment seizure risk under a new CPSC filing rule. Independent Seller Shipping was discontinued on February 25, 2026. The policy applies to all United States orders and to third-party fulfilment providers, Amazon Multi-Channel Fulfilment operations included.

The review-bait clause is the one that connects to the week's other stories. A rating on a marketplace listing functions as a purchasing signal, and it feeds ranking, advertising eligibility and conversion. A printed card asking for five stars is the cheapest possible way to move that signal without moving the underlying product quality. Banning the card does not stop the practice, since sellers have email, packaging design and follow-up messaging. It does establish that the platform now treats an inserted review request as a packaging violation rather than a marketing choice, which changes the category of the penalty and the system that applies it.

A 36 percent reach figure, and the audit that arrives in 2027

DAZN Media+ announced a measurement partnership with AudienceProject in Sydney on September 3, 2026, and PPC Land covered it on September 7.

DAZN Media+ is the advertising sales arm for Foxtel, Kayo Sports and BINGE. AudienceProject is a Copenhagen-based independent measurement company. The arrangement combines census-level campaign delivery data from DAZN Media+ with AudienceProject's own audience panel, producing demographics, co-viewing, reach and frequency, and cross-platform overlap on a single methodology. The stated purpose is incremental reach: distinguishing viewers reached through Kayo Sports and BINGE beyond linear television from viewers already counted on linear.

The number in the announcement is the part that requires care. DAZN Media+ cited internal testing showing streaming added 36 percent reach to Australian Football League campaigns and 27 percent to National Rugby League campaigns. Those are vendor-supplied figures. No sample size, measurement window or methodology was disclosed alongside them.

The announcement effectively concedes the point. Nev Hasan, chief sales officer at DAZN Media+, framed the AudienceProject arrangement as enabling the company for the first time to independently demonstrate reach across its platforms. A seller stating that independent demonstration begins now is also stating what the preceding figures were, and the 36 percent number quoted in the same release belongs to the earlier category.

The timing compounds it. The measurement solution becomes available in 2027, with no month specified and no pilot phase described. Australian campaigns running through the spring carnival, the summer cricket season and the January tennis window will be planned and traded on existing approaches. The independent number arrives after the buying decisions it would have informed.

There is also an internal inconsistency the release does not resolve. The body text lists Kayo Sports, BINGE, Foxtel Go and linear television as the covered platforms. Hasan's quote additionally names Foxtel Now delivering audiences alongside linear television. Which properties are in scope is therefore ambiguous in the announcement itself.

Context makes the gap matter more rather than less. Australian video advertising reached a record 4.9 billion Australian dollars in the first quarter of 2026, up 15.3 percent year on year, with video accounting for 73.2 percent of display spending. A May 2026 IAB Australia survey found only 63 percent of respondents consistently obtaining adequate data from ad-supported subscription services. That is a market where a large and growing share of spend is transacted against signals a majority of buyers describe as adequate and a substantial minority do not.

The structural distinction worth drawing is between this arrangement and the Video Futures Collective collaboration announced on August 24, 2026. The DAZN Media+ contract is a commercial agreement covering one seller's portfolio, producing deduplication within that portfolio. The Video Futures Collective is an industry-level initiative pursuing comparable numbers across competing platforms. A seller measuring its own overlap solves the seller's problem. Whether the two efforts converge on a common currency, and on what timetable, is not addressed by either announcement. Panel size, demographic reporting capabilities, local audit arrangements and pricing were all left undisclosed.

Also noted

September 7: Morrill Creative LLC, founded by Kelly Morrill, launched Kids Podcasts, a free iPhone application putting a curated library of more than 200 hand-vetted shows for ages 2 to 12 behind a parent PIN, with no open search, no algorithmic recommendation, no advertising and no third-party SDKs, in a children's audio category where poor measurement is estimated to waste 590,000 dollars of every 1 million spent. PPC Land
September 6: Seven researchers from Taobao and Tmall Group reported a 0.36 percent GMV increase in a 41-day online A/B test using Direct Causal Effect Optimization, which replaced a single hand-tuned fusion formula with request-specific weights across 17 funnel predictions, while conceding in arXiv 2608.25635v1 that the critic was trained on observational logs and assumes unobserved confounding is absent. PPC Land
September 5: Safari dropped its legacy JavaScript module loader to fix top-level await behaviour, a browser engine change that alters execution order for any advertising or measurement tag relying on module loading. PPC Land
September 5: Amazon replaced one-at-a-time regulatory filing with a bulk Toy Restriction Framework compliance tool, cutting the per-item administrative cost for sellers facing product safety documentation requirements across large catalogues. PPC Land
September 7: Google confirmed that prices shown through Search booking partners are customised rather than fixed, meaning the figure a user sees in a booking result is not necessarily the figure another user sees for the same inventory. Search Engine Roundtable