Rival search engines and AI chatbots seeking access to Google's European search logs will receive a heavily filtered version of them, according to Damien Desfontaines, the privacy engineer who worked with the European Commission on the specification and who today published a step-by-step account of the anonymisation method Alphabet must apply before any record changes hands.

In Short

Google has to give some of its European search records to competing search engines, and the privacy expert who helped the European Commission write the rules today explained how those records get scrubbed first. This affects everyone who searches on Google in Europe, the rival search engines and AI chatbots that want the data, and the advertisers and publishers whose traffic depends on how those rivals rank pages. Before anything is handed over, Google must remove who searched, delete searches containing rare words or very long text, and blur locations until every record blends in with at least 1,000 people, while buyers face audits and a ban on using the data for ads.

A search log worth fighting over

Google Search has held around 90 percent of the European search market for more than a decade, according to Desfontaines, who writes under the name Ted on his blog Ted is writing things. The Commission's own guidance on the case describes a share above 90 percent sustained for decades. Market share on that scale produces something competitors cannot buy elsewhere: a record of what hundreds of millions of people type, what they click, how long they stay and when they return.

Desfontaines illustrates the value with one of his own sessions. He typed "grand prix", stayed on the results page for under 2 seconds without clicking, then searched "grand prix attack", spent 1.5 seconds on a chess.com page, 37.8 seconds on a Reddit thread, and finally left for a chess video without returning. From that sequence alone, an engine learns that the phrase can mean a chess opening rather than Formula 1, that a well-branded result was not especially useful, and that video suits the topic. Scaled up across a user base like Google's, he writes, "this data is absolute gold to improve a search engine."

The volume gap is not a matter of engineering skill. According to the post, 15 percent of the queries Google Search sees each day are brand new, a figure Google publishes itself, and no amount of pipeline optimisation compensates for having orders of magnitude less data. Google's own scientists described the phenomenon in 2009 as the "unreasonable effectiveness of data." Across the Atlantic, courts have measured the same asymmetry: users enter nine times more queries on Google than on all rivals combined, rising to 19 times on mobile, according to findings from the US search monopoly case.

Two years and a dataset nobody used

The Digital Markets Act, passed in 2022, addresses that gap directly. Article 6(11), quoted in the post, states that the gatekeeper "shall provide to any third-party undertaking providing online search engines, at its request, with access on fair, reasonable and non-discriminatory terms to ranking, query, click and view data in relation to free and paid search generated by end users on its online search engines." The same article requires that any such data constituting personal data "shall be anonymised."

Google began offering competitors access to an anonymised dataset in March 2024, the month DMA obligations became legally binding. Uptake was negligible. According to the Commission, Alphabet's initial offer removed between 90 and 100 percent of unique search queries and excluded AI chatbots from the pool of eligible recipients, a finding PPC Land reported when the binding decisions were adopted. A counterproposal published on TechPolicy.Press on April 28, 2026 estimated that Google's dataset omitted more than 99 percent of distinct queries and about 42 percent of total query volume.

Desfontaines is blunt about the cause. After the Commission asked for his help in October 2025, he writes, he concluded that "the anonymization strategy chosen by Google was the single greatest issue with their compliance approach." His assessment of the result: "Their design choices rendered the data all but useless for optimizing a search engine." In a footnote, he adds that given Google's internal expertise, "it's pretty clear that they could have done much better if they wanted to."

The Commission opened specification proceedings on January 27, 2026. Eighteen European industry groups wrote to Commission President Ursula von der Leyen in March demanding action on Google's search compliance, preliminary measures followed on April 16 with a consultation running to May 1, and the Commission adopted the binding decision on July 16, 2026.

One timing detail in the post does not match the public record. Desfontaines writes that the proceedings ended "last month," yet the July 16 adoption date falls nearly two months before the post's September 10 publication. His statement that the full decision was published "this week" fits the Commission's July guidance, which said the complete technical description would follow later. The full decision itself was not among the documents reviewed for this article.

The adviser and his disclosures

The account comes from a participant rather than an observer, and Desfontaines states as much. His personal page lists him as founder of Hiding Nemo, an independent consultancy working with privacy-enhancing technology, with clients including the European Commission, Vinted, The Everyone Project, TikTok and Oblivious. "I previously led the anonymization consulting team at Google," the page states. He later worked as a staff scientist at Tumult Labs, whose open-source differential privacy library was transferred to OpenDP after LinkedIn acquired the company, and he obtained his PhD at ETH Zürich in 2020 under David Basin.

His role in the case spanned analysis of Google's proposal, written questions and meetings with Google and its competitors, the design of an alternative method, data analysis of its effectiveness, and responses to consultation feedback and Google's technical arguments. The post carries a disclaimer: "This blog post is not part of our contract; it is strictly based on publicly available information, and I am not being paid for it." The Commission reviewed the text for factual accuracy and confidentiality, according to Desfontaines, while the opinions remain his own. Readers weighing his characterisation of Google's conduct have that context: a paid adviser to one side, describing the other side's proposal in sharply critical terms.

What "anonymised" means in this case

The legal bar comes from Recital 61 of the DMA, which will be familiar to anyone who knows Recital 26 of the GDPR. Data is anonymised, it reads, "if personal data is irreversibly altered in such a way that information does not relate to an identified or identifiable natural person." Draft Joint Guidelines from the European Data Protection Board and the Commission then add two features that Desfontaines calls fairly unique.

First, only the person running the search is protected. If Alice searches for Bob's full name and a result shows Camille's personal website, only Alice's identity must be hidden; the query and results can still carry Bob's and Camille's personal information. The consequence is unusual. The shared dataset still contains personal data, so the GDPR continues to apply to recipients, whereas most organisations anonymise data precisely to take it out of GDPR scope. The case-specific approach runs alongside the European Data Protection Board's general Guidelines 02/2026 on Anonymisation, open for public comment until October 30, 2026.

Second, anonymisation can be achieved through a combination of technical measures, which alter the data, and non-technical measures, which cover everything else, from contracts to access controls and logging. Technical measures must be prominent but may leave a non-zero residual risk, provided the non-technical layer mitigates it.

The team turned "residual" into three tests. Risk exceeds that level if an adversary can pick a specific person and, with significant probability, retrieve some of their search records with high certainty; if an adversary can pick a random record and, with significant probability, re-identify the user with high certainty; or if an adversary can re-identify a large number of users with high certainty using reasonably simple methods. The threats in view were curious or indiscreet employees, often called insider risk, and product or analytics teams at recipient companies breaching their contracts to enrich other datasets.

Formal guarantees were ruled out. "We're not trying to provide formal guarantees that hold for any attacker, like in differential privacy; this would be largely incompatible with achieving reasonable utility for this use case," Desfontaines writes. The design target was realistic adversaries operating under extensive contractual constraints.

Utility carried legal weight too, because Article 6(11) requires choosing the method that retains the most value among those meeting the privacy bar. Discussions with search providers pointed to three areas where the gap hurts most: crawling and indexing, query understanding, and ranking and quality. All three depend on the long tail. Rivals do not need Google's logs to answer "capital of switzerland"; they need them for novel, ambiguous queries whose best answers sit on rarely visited sites, and for context such as which results a user skipped.

Five filters before a record leaves Google

The specified scheme combines eligibility rules, technical measures and non-technical measures. According to Desfontaines, the technical procedure has five main stages, described fully in paragraphs 594 to 611 of the decision. The Commission's public guidance groups them into three steps.

Stripping the person out

The first stage is de-identification. User identifiers, precise timestamps and other attributes that could act as a fingerprint are removed from each record; the Commission's guidance cites Google usernames, IP addresses, advanced filters and input format as examples. The result is what Desfontaines calls a haystack: loose individual queries that cannot be chained into a search history. He counts it as the most important privacy step and a steep utility cost, since recipients never see full sessions.

Removing rare words and entities

Each query passes through a personal data detector that extracts full names, addresses, phone numbers and location coordinates, and each separate word is extracted as well. If any of those entities was searched by fewer than 50 distinct users in the past year, the entire record is deleted. The target is secrets: package tracking numbers, links to nonpublic URLs, passport numbers. Typos get swept up too, false positives that Desfontaines calls unavoidable. The April preliminary version framed the same idea as a weekly allowlist built from 13 months of signed-in European search data, with an entity admitted for five years once more than 50 unique signed-in users had searched it.

Cutting unusually long queries

Every query longer, in characters, than the 95th percentile for its language is removed. For English, the threshold falls somewhere between 50 and 75 characters, according to the post. Long strings are more likely to be accidental pastes of private documents and to carry identifying detail, though the cut also discards innocuous long-tail behaviour.

Coarsening location and device

User location is replaced by a NUTS 3 region, an administrative unit designed to hold between 150,000 and 800,000 inhabitants, and device type is reduced to desktop, mobile or tablet. The procedure then checks that at least 1,000 signed-in users share the same region, device category and inferred language. If not, location drops to country level. If the threshold still is not met, the record goes. Desfontaines notes that this is k-anonymity at work, and that 1,000 is a hard floor: 95 percent of users sit in groups of 25,000 or more, a figure the Commission also published.

The preliminary measures had been considerably finer. In April, location was to be expressed through S2 map cells containing at least 1,000 signed-in users and covering at least 3 square kilometres, with further generalisation triggered only when fewer than 50 signed-in users shared the same combination of language, location and device.

Rebuilding short refinement chains

The final stage, mini-sessionisation, restores a little of what de-identification destroys. Records are re-linked where a user clicked something on the results page that generated a second query, such as a "Did you mean" suggestion or an entity chip, or typed a new query that contains the first as a substring. Chains are capped at three successive queries. The conservative trigger means a mini-session typically discloses nothing beyond its final query, according to the post, while giving rivals evidence of how people rescue vague searches.

None of it is presented as airtight. "But nobody is arguing that these technical measures bring the risk to zero!" Desfontaines writes. Their job is to reach the residual level; the contracts do the rest.

Google's two-hour claim and what it changed

Google has fought the method in public. Kent Walker, president of global affairs at Google and Alphabet, argued on July 16 that the search order would expose Europeans' private searches to unfamiliar companies without adequate anonymisation. Earlier, a letter to the Commission from Sergei Vassilvitskii, a distinguished scientist at Google, said the company's team had re-identified individual users from data processed under the proposed method in under two hours, according to a May 13, 2026 post on the Online Safety newsletter summarising Reuters reporting.

Desfontaines disputes the weight of that result. Google, he writes, "ignored non-technical measures" and "threw more resources at the problem than would ever be realistic for a data recipient," based its claims on the preliminary measures, and largely neither validated them nor quantified its certainty. The two sides also tested against different standards. The Commission's team measured attacker success against specific individuals, random records and re-identification at scale; Google hunted for worst-case edge records. Is a handful of re-identifiable records proof of failure? Under the residual-risk framework, not necessarily, and Desfontaines says most of Google's findings fit within it.

Some did not, and the measures moved. According to the post, the team excluded all sponsored results from the dataset, adopted much larger geographic regions and raised the metadata threshold. The last two changes appear to correspond to the shift from S2 cells and a 50-user combination check to NUTS 3 regions and a 1,000-user floor.

The sponsored-results change deserves care. Article 6(11) explicitly covers data on "free and paid search." The April preliminary measures kept interaction and ranking data for paid results at the level of the ad block while excluding individual ad URLs. Desfontaines now describes all sponsored results as excluded, while the Commission's July guidance refers to URLs for paid results being removed. The two descriptions differ on whether any block-level paid data survives, and the documents reviewed do not resolve it.

Two footnotes carry allegations likely to feature in any court challenge. Desfontaines writes that "several respondents were remunerated by Google for their participation in the public consultation," citing paragraph 47 of the decision. He also points to paragraphs 717 and 718, and footnote 1033, for what he calls surprising delays and arbitrary technical limitations in the way Google shared the data sample used for testing.

Who can buy, and under what rules

Eligibility is the first filter. An applicant must provide online search services in the EU and either have done so for at least two consecutive years or, if founded more recently, have raised more than 50 million euros in capital investment. It also needs at least 50,000 average monthly users in the EU over the past year. Alphabet may reject companies under EU restrictive measures or controlled by a non-EEA state actor posing a "serious and structural" data protection risk, a test set out in paragraph 128(b). Under the Commission's guidance, recipients must process the data in the EEA or guarantee essentially equivalent protection under GDPR Chapter V, Alphabet may assess security risks before sharing, and the Commission can exempt Google from specific recipients on public security grounds under Article 10. AI chatbots with a web search function count as online search engines.

The non-technical measures, set out in paragraphs 818 to 828, are strict. Recipients must process the data in a segregated environment and may not copy it elsewhere or pass it on. Use is limited to optimising search; chatbot operators may apply it to search and retrieval but not to training or fine-tuning general-purpose models, and the Commission's guidance separately bars use for advertising, consumer profiling or systematically replicating Google's results. Linking with other datasets, attempting re-identification and augmenting data to reverse the technical measures are all prohibited, and the attempt alone is a breach whether or not it succeeds. Data must be deleted 13 months after receipt. Every access must be logged, and models trained on the data must be evaluated to confirm they do not reproduce re-identifiable records.

Audits sit on top. An independent auditor must confirm a recipient's safeguards before any data flows, a first compliance audit follows within six months of processing, and annual audits continue after that, with ad hoc audits available to the Commission.

Sharing parameters are fixed as well. According to the Commission, Alphabet must strip invalid traffic from the dataset, deliver data with a latency of at least seven days, and provide access for up to five years per beneficiary. Precise interaction durations are replaced by time intervals; the April draft specified six bins running from 0-10 seconds up to more than 20 minutes. Pricing is limited to Alphabet's incremental costs of preparing, storing and transmitting the data plus a return capped at its weighted average cost of capital. An extra margin, capped at the operating margin of Google Search, is possible in exceptional cases, including for beneficiaries operating at gatekeeper scale, but never for micro, small or medium-sized enterprises. Prospective buyers can test three samples first: a small real-data sample, a synthetic dataset and a larger representative dataset, the last available only to applicants holding an auditor's reasonable assurance report.

Google has already opened the door procedurally. Its European Search Dataset Licensing Program, documented by Search Engine Roundtable and reported by PPC Land on September 1, takes applications through a Google Form.

Deadlines through January 2027

The Commission's milestones run in sequence from the July adoption. By the end of August 2026, Alphabet had to submit its eligibility application form for review and publish a webpage explaining how to apply. By September 2026, it must provide template licence agreements, the test samples, and its cost items and estimates. By November 2026, the anonymised dataset must be finalised, along with technical information on latency and personal data detectors. By January 2027, the pricing offer must be complete. Desfontaines summarises the window as six months to implement the technical measures, with competitors potentially receiving data in early 2027.

The measures face biennial reviews, and Article 8(9) lets the Commission reopen proceedings if facts change, including through independent third-party testing of the anonymisation. Desfontaines does not expect a quiet implementation. "I expect that Google will continue resisting sharing valuable data with competitors in any way they can," he writes, naming the courts and lobbying aimed at the media, regulatory bodies and professional communities as likely venues.

Why the marketing industry is watching

For search marketers, the direct effect is on competition rather than on campaigns. The Commission argues that more credible search alternatives mean more traffic sources for businesses and more competition in the advertising market. Whether that happens depends on who buys, and how quickly they can turn a filtered, week-old log into better rankings.

Advertising data stays largely outside the arrangement on both sides of the Atlantic. The European scheme strips paid result URLs and bans advertising uses outright. In the United States, Judge Amit Mehta's September 2, 2025 ruling likewise required Google to share search index and user-interaction data, though not ads data, under a six-year remedy that Google appealed in January 2026.

Publishers have a direct stake in the AI dimension. Retrieval for grounding, the step in which a chatbot fetches web pages before composing an answer, is an approved use, and the long-tail click signals in the dataset bear directly on which rarely visited pages rivals decide to crawl and index, and by extension which pages their retrieval systems can surface. One of the most widely used chatbots with a search function is already subject to separate EU rules: the Commission designated ChatGPT a very large online search engine on August 31, 2026, after OpenAI declared 159.1 million average monthly users in the EU. Any chatbot operator would still have to pass the eligibility checks and audits, accept the ban on training general-purpose models with the data, and, if it met the gatekeeper thresholds, face the additional pricing margin.

Google's privacy objections have also been set against its own data purchasing. On August 14, 2026, the company was selected as the winning bidder, at 10 million dollars, for deidentified data from bankrupt Spirit Airlines, a purchase that Alan Chapell contrasted with Google's posture in its own litigation, where the company has objected to sharing search data on grounds that effectively treat genuine anonymisation as close to impossible.

The search data rules arrive amid wider enforcement. A week after the July decisions, the Commission fined Google 890 million euros over self-preferencing in Search and steering restrictions in Google Play, and Google said it was weighing an appeal ahead of a September 21, 2026 compliance deadline. Specification decisions such as the search data order carry no fine, but they bind Google to a technical design that, according to Desfontaines, will be reviewed and contested for some time. This is, as he puts it, "probably not the last you hear of this!"

Timeline

Summary

Who: Damien Desfontaines, founder of the consultancy Hiding Nemo and former lead of Google's anonymisation consulting team, who worked with the European Commission on the case under paid engagements; Alphabet and Google as the designated gatekeeper; rival search engines and AI chatbots with search functions as potential recipients; and the advertisers and publishers whose traffic depends on those services.

What: A detailed public account of the anonymisation scheme Google must apply before sharing European search data under Article 6(11) of the Digital Markets Act: de-identification, removal of records containing entities searched by fewer than 50 distinct users in a year, removal of queries above the 95th length percentile, NUTS 3 location generalisation with a 1,000-user floor, and mini-sessions of up to three queries, backed by eligibility rules, contractual limits, 13-month retention and yearly audits.

When: Desfontaines published the account today, September 10, 2026. The Commission adopted the decision on July 16, 2026, after opening proceedings on January 27, 2026. Alphabet must finalise the dataset by November 2026 and its pricing by January 2027, with competitors potentially receiving data in early 2027.

Where: The European Union and European Economic Area, covering Google Search activity by European users, with recipients required to process data in the EEA or under GDPR-equivalent transfer protections.

Why: Google's March 2024 dataset attracted no meaningful uptake after its anonymisation removed most unique queries, according to the Commission. The specification sets out a method intended to meet the DMA's anonymisation standard while preserving the long-tail query, click and ranking data that competitors say they need to improve crawling, query understanding and ranking, including for AI search products.