AI Search
Google's August 2026 spam update reached every language on August 18, as Apple rewrote its EU App Store fee terms and ChatGPT ad units swapped advertisers.
Two weeks of unexplained ranking movement ended with an announcement. At 09:27 Pacific on Tuesday, August 18, Google logged the August 2026 spam update on its Search Status Dashboard, and the entry carried the two qualifiers that matter most to anyone tracking visibility across borders: it applies globally, and it applies to all languages. The rollout may take a few days.
That was the whole of the official communication. What followed across the approved trade press on August 18 and 19 was a set of stories that, taken together, describe a market in which the rules of visibility are being rewritten in three places at once. Google is enforcing its own. Apple is renegotiating the price of distribution in Europe under regulatory pressure. And the AI answer layer, where a growing share of brand discovery now happens, turns out to be selecting brands before it searches at all.
The August 2026 spam update is the third of the year, following one released on March 24 at 12:18 Pacific and another on June 24 at 09:03 Pacific. The March release closed in roughly 19 and a half hours. The June release closed on June 26, about 48 hours after it began. Google’s own guidance for this one is looser: a few days, with no completion date named.
Barry Schwartz logged the announcement at around 12:30 Eastern and noted the distinction that tends to get lost in the first day of commentary. Spam updates of this kind target violations of the general search spam policies. They are not the link spam update, and they are not the site reputation abuse enforcement that has run on its own track since September 2024. Google declined to quantify how many queries are affected, which is standard.
On LinkedIn, Google Search Central described it as a normal spam update. The phrasing is doing work. Spam updates refresh the classifiers that sit behind SpamBrain, the automated detection system Google has credited with the bulk of its spam suppression since 2022, and they re-apply existing policy rather than introduce new policy. No new policy accompanied the August release. The categories in force were already on the books: expired domain abuse and scaled content abuse, both added in March 2024; site reputation abuse from September 2024; back button hijacking, added in April 2026 and enforced from June 15; and the AI Overview and AI Mode violations added on May 15, 2026. Older categories run underneath all of it, among them cloaking, inauthentic mentions, link spam, and doorway pages, the practice of building near-identical pages for slight keyword variations that funnel traffic to a single destination.
The timing matters because the update did not arrive into calm conditions. Tracking tools registered volatility on August 1 through 3, a further unconfirmed spike on August 5, and another period on August 12 and 13. That last window overlapped with a separate incident: Search Console performance reports lost impressions and clicks starting around August 12, a logging failure rather than a ranking event. Sites that were reading their August data for evidence of an unannounced update were, in part, reading a broken meter.
Recovery expectations are worth stating plainly, because they differ from core updates. Google’s documentation says sites in violation may rank lower or not appear at all, and its guidance on returning has consistently been framed in months rather than days, with no guarantee attached. There is a specific asymmetry in the link spam case: once spammy links are neutralised, the ranking benefit they conferred does not come back, so recovery restores the site to what it would have earned without them, not to its prior position. Across the 173 confirmed Google updates recorded between 2000 and 2026, 52 have been core updates and 26 have been spam updates. Core updates average around 17 days. Spam updates have ranged from under a day to the 27-day August 2025 rollout, which SISTRIX later characterised as producing minimal visible ranking movement despite its length.
Around the update, smaller mechanical changes accumulated on the same two days. Google confirmed that Search will eventually support the HTTP QUERY method, a proposed standard that allows a request to carry a body while remaining safe and idempotent, which has implications for how large parameterised requests are crawled. The Related searches for module appeared with underlined hyperlinks in live results. And Search Profiles, the creator-facing surface whose follower threshold dropped to 35,000 on August 13, gained the ability to edit a handle after creation.
A study circulating through the search community this week makes an argument that cuts against most of what has been written about optimising for AI answers. Suganthan Mohanadasan, a Norwegian search consultant based in Dubai and co-founder of Keyword Insights and Snippet Digital, captured raw network traffic from ChatGPT conversations on July 24 and 25, 2026, and examined the JSON response key that carries the model’s own generated search queries. The key had been renamed from search_model_queries to search_queries in early August 2026. His finding, reported by PPC Land on August 18, is a 33-fold gap.
Brands that appeared in ChatGPT’s first self-generated search query reached the final answer 68.9 percent of the time. Brands that were retrieved during the session but never named in any query reached the answer 2.1 percent of the time. The sample is small and should be read as such: 60 conversations overall, 57 used for citation measurement, 27 for the first-query test, built from 12 fresh category queries plus three repeat runs, all from a single logged-in ChatGPT Plus account in Dubai.
The mechanism is what makes it interesting. In 21 of the 27 first queries, the model injected brand names that the user had never mentioned. Those names came from the model’s parametric memory, formed during training, before any page was fetched. Retrieval then ran against a query that already contained the answer’s shortlist. Of 3,554 pages retrieved across the sessions, 110 earned a citation, a rate of 3.1 percent. Citation probability also decayed sharply by position within a domain group: 5.2 percent for the first result, 4.6 percent for the second, 2.4 percent for the third, 1.7 percent for the fourth, 0.6 percent for the fifth, and 0.3 percent from the sixth onward.
Mohanadasan revised his own headline on August 17 because the original, he said, could be read as claiming the outcome is decided before the search runs. The distinction he draws is narrower and more defensible: injection into the query is a strong predictor of inclusion, not a guarantee of it. Rand Fishkin, commenting on LinkedIn, put the practical consequence of the finding more bluntly, describing conventional AI visibility strategies as something that might sit somewhere between potentially misleading and totally useless.
The study lands in the same week as a smaller observation about the paid layer of the same interface. Brodie Clark of SERP Alerts documented ChatGPT ad units loading one advertiser and then rapidly switching to another before the user has seen the first, and asked the question that follows from it: on a cost-per-impression model, who pays for the impression that flickered. Schwartz, writing it up on August 18, speculated that an ad server deficiency is the likeliest explanation but did not claim to know the cause. AdExchanger’s August 19 roundup filed the same behaviour under the heading of ad hallucinations, alongside OpenAI’s expansion of ChatGPT advertising into Mexico and Brazil and a European rollout planned before the end of the month.
Two findings, one organic and one paid, point at the same structural fact. The surface where an increasing share of discovery happens does not yet behave like a stable, auditable inventory. Retrieval favours what the model already knows. Impressions are not reliably what they appear to be. Both are measurement problems before they are anything else.
Apple published a rewrite of its European App Store terms on August 18, and the headline change is the removal of a fee that had been a specific irritant to ad-funded businesses. The Core Technology Fee, charged per install to developers operating at what Apple called extraordinary scale under the Alternative Terms Addendum, is replaced by a 5 percent Core Technology Commission on sales made through alternative distribution and web distribution. The new terms take effect on October 1, 2026, superseding the existing Alternative Terms and the StoreKit External Purchase Link Entitlement through Attachment 14 of the Developer Program License Agreement.
The difference between a per-install charge and a percentage of sales is the difference between a fixed cost and a variable one, and it falls unevenly. A free, advertising-supported app generating large install volumes with no digital sales carried real exposure under the old structure and carries none under the new one. A paid or subscription app with high revenue per install sees the reverse.
The rest of the schedule moves in the same direction of lower headline rates with more categories. Apple In-App Purchase drops from 30 percent to 26 percent, with a 15 percent reduced tier covering the Small Business Program, Mini Apps, the Video Partner Program, and auto-renewable subscriptions after the first year. Alternative payment processing inside an app is charged at 20 percent, or 10 percent at the reduced tier and for second-year subscription renewals. Offers linked out of the app carry a 15 percent store services commission, 10 percent reduced, and only sales completed within seven days of the link tap qualify. The alternative payments entitlement now covers 29 regions: the 27 EU member states plus Iceland and Norway.
Two operational constraints deserve attention because they set planning horizons. Developers must hold their chosen payment configuration across every EU storefront for 12 months, with no mid-year switching between Apple’s own processing and an alternative. And monthly reporting is due within 15 days of month end, covering refunds, corrections, renewals, one-time purchases, and failed transactions.
Marketplace eligibility also widened. Where operating an alternative app marketplace previously required EU establishment, seven qualification routes now exist, including a Dun and Bradstreet low or below-average risk rating, a listing on a World Federation of Exchanges member or Euronext, venture funding from specified lists, a recent unqualified financial audit, nonprofit or educational or government status, a 1 million dollar standby letter of credit, or one million first-year iOS installs globally in the prior calendar year.
The regulatory backdrop explains the timing. The European Commission fined Apple 500 million euros on April 23, 2025 over anti-steering restrictions. On July 8, 2026, the General Court dismissed Apple’s challenge to its gatekeeper designation under the Digital Markets Act. Apple’s own framing of the August 18 package is that it resolves the company’s disagreements with the Commission, which is a statement of intent rather than a confirmed outcome; the Commission has not closed the matter.
On August 14, a virtual auction in the United States Bankruptcy Court for the Southern District of New York selected Google as the winning bidder for the business data of Spirit Airlines. The price was 10 million dollars. The alternate bidder, Mercor.io Corporation, offered 7.5 million. The case, 25-11897 before Judge Sean H. Lane, has been running since Spirit filed for Chapter 11 on August 29, 2025; bidding procedures were approved on June 22, 2026, objections were due at 4:00 pm Eastern on August 17, and the sale hearing was set for 11:00 am Eastern on August 19.
The asset schedule is unusually specific. It covers roughly 100 million emails across 80,000 Microsoft 365 accounts, 516 source code repositories containing around 30 million lines of code, 7.5 billion transactions drawn from the Navitaire revenue system, 175,658 employee records reaching back to August 1986, records of 763,391 flights from April 2023 onward, and the finance, legal, and operations archives.
What is excluded is equally specific, and it is the part a marketer would recognise. The sale leaves out 97.5 million customer profiles, 50.2 million Free Spirit loyalty members, marketing email lists containing 13.7 million active addresses, more than 15.7 million contact centre chat sessions, and website analytics. In other words, the consumer marketing database, the part with the clearest conventional commercial value, is not what changed hands. The operational exhaust is.
Alan Chapell noted the tension between the purchase and Google’s posture in its own litigation, where the company has objected to providing search data to rivals on grounds that effectively assert genuine anonymisation is close to impossible. Matt Mickiewicz framed it on X as a measure of scarcity: AI labs are sufficiently short of fresh training data that the emails and documents of a bankrupt airline reach eight figures.
The AdExchanger roundup on August 19 placed the transaction in a category rather than treating it as a one-off, grouping it with reporting that OpenAI, Anthropic, and other model operators are buying failed startup logs, internal messages, and Slack archives, and with a 404 Media investigation that used AirTag tracking to follow rare books to an Amazon facility where they were despined, scanned, and discarded. The common thread is that the training-data market has moved past the public web and into the corporate estate, including the estates of companies that no longer exist.
The Association of National Advertisers published a report titled Retail Media Measurement Standardization on August 18, built from consultation with more than 40 heads of media at brands including Mondelez, PepsiCo, Hershey, Colgate Palmolive, Clorox, and Intel, and with retail media networks including Walmart Connect, Sam’s Club Connect, CVS Media Edge, Instacart, Albertsons Media Collective, BJ’s Media Edge, and Roundel. Amazon, the largest retail media seller by some distance, did not participate.
The recommendations reported by Digiday are procedural rather than technical: valid third-party measurement and validation, transparent disclosure of measurement methods and logic and definitions, a standardised 14-day loopback window for attribution, performance context that extends beyond attributed sales, and a single agreed definition of the word outcomes. Jackson Bazley, the ANA’s EVP of measurement, described the current state of cross-network comparison as evaluating purple to bananas to 9.7 stars, and said the response from the industry had been that everything in the document makes sense and nothing is out of left field. Liz Roche, VP at Albertsons Media Collective, said standardisation is realistic but has to start with transparency. The work builds on a 2024 MRC and IAB report that set baseline metrics. The named third-party validators are IAS and DoubleVerify.
The absence of an agreed measurement layer is the same constraint that surfaced in a separate Digiday piece on August 19, which found a brand-side buyer declining to adopt automated buying agents for exactly that reason. Paras Shah, senior director of digital media at Georgia-Pacific, framed it as a sequencing problem: an agent will only automate a repetitive task, and if there is waste in the system, the automation preserves the waste. He did not rule the technology out, saying the company would consider buying agents in future but treats getting the best quality impression at the lowest possible cost as the more important objective. Ian Maxwell, CEO of Converge, described a tendency to confuse automation with intelligence, and said agentic buying will not improve media performance if what it automates is inefficient workflow, poor data, or bad commercial decisions. Mateusz Jedrocha, chief product officer at Adlook, said the industry is thinking about integration in the wrong order.
That scepticism has a supply-side counterpart. AdExchanger profiled Tuple on August 18, a demand-side platform that went live in March 2026 with eight customers and a deliberately narrow footprint: one SSP integration at present, Media.net, and a stated ceiling of five, with three described as the target. Founder Doug Lauretano spent a decade on the supply side at OpenX and Media.net before building deal ID and curation products at CivicScience from 2021, and started on the publisher side at Fortune, the Wall Street Journal, and CNN Money in the early 2000s. His argument is that a platform with minimal duplication has the best chance of performing well, and that opacity was the original sin of legacy ad tech.
Three stories, three constituencies, one recurring position: the measurement problem is prior to the automation problem. Standardising retail media outcomes, refusing to layer agents on unreliable inventory, and building a DSP that connects to three supply paths rather than thirty are all versions of the same argument about knowing what was bought before optimising how it is bought.
Every morning, get a curated digest of the last 24 hours in programmatic advertising
