The State of Search

Can I block AI training without losing search traffic?

settled The record agrees. A change here would itself be news.

With Google and OpenAI, yes — both state in their own documentation that the training and search crawlers are independent. Anthropic runs separable crawlers but publishes no equivalent guarantee. In every case the block is a request rather than a lock — and on Cloudflare-proxied sites, the layer enforcing it can undo the operator's own separation: taking Cloudflare's September 15 default blocks Google's, Apple's and Bing's search crawlers too, because those crawlers are multi-purpose.

Verified against the record 2026-09-16 · record last moved September 16, 2026

The three operators the question actually turns on each run a search crawler and a training crawler that can be addressed separately in robots.txt. Two of them say in public what blocking one does to the other; the third does not.

Compliance is a separate question from policy. TollBit measured roughly 15% of the AI page fetchers it tracked in the first half of 2026 reaching European publisher URLs marked disallowed, with OpenAI's ChatGPT-User among the bots most often crossing blocks aimed specifically at it. No operator publishes a compliance rate against its own declared behavior, so a block is worth logging against rather than trusting.

Since August 21, 2026 a fourth party sits in this decision for Cloudflare-proxied sites. Bot Preference Sync conditions continued access for crawlers that both search and train on four public disclosures, which converts what was an operator's private practice into something a publisher can check.

The standard underneath any of this is itself still being drafted. Revision 08 of the IETF's AI Preference Vocabulary, published September 16, 2026 and co-authored by Google's Martin Thomson, split its single "AI Training" section into separate "AI Training" and "AI Use" sections while keeping "Search" as its own category — giving sites a more precise vocabulary to state the training-versus-search distinction this claim already tracks operator by operator. The draft expires March 18, 2027 unless renewed, and no operator has said whether it will implement the new categories.

aeo · engines · law-money

Is AI-crawler access now a paywall rather than a block?

moving Actively changing — a rollout, a deadline, or a live fight.

On the Cloudflare-proxied publishers the wire tracks, yes for some — access is starting to split by whether the crawler's operator pays rather than by whether the site allows AI crawling at all. Four large publishers now block OpenAI's GPTBot outright while continuing to bill Anthropic's ClaudeBot and Perplexity's crawler through Cloudflare's Pay Per Crawl. Other large sites reject the split entirely and block every crawler, paying or not.

Verified against the record 2026-09-16 · record last moved September 16, 2026

A different kind of payment surfaced September 16, on the answer side rather than the crawler side. Google confirmed an early-stage pilot that pays a small set of publishers when their content "contributes significantly" to an answer in Gemini, AI Overviews or AI Mode — content that only confirms a fact or shows up as a post-answer link doesn't qualify. Participants see a monthly earnings figure inside a Search Console panel and can opt out any time; Google hasn't disclosed a rate or eligibility criteria, and sources described the earnings calculation as "quite black box" with early payouts small next to ad revenue. It meters something different from the Cloudflare split above: Pay Per Crawl charges for access before an answer is generated, this pays for contribution after one is.

The wire runs its own crawler access panel, probing 55 publisher domains daily as GPTBot, ClaudeBot, PerplexityBot and a browser user-agent. Between the September 12 and 13 captures, three People Inc. properties (Allrecipes, Investopedia, People) and The Atlantic made an identical change: GPTBot and the browser probe went from a 200 to a Cloudflare-managed 403, while ClaudeBot and PerplexityBot kept drawing the same 402 Pay Per Crawl response as before — the same site charging two crawlers for the access it now refuses a third.

That split sits on top of Cloudflare's Pay Per Use mechanism, announced July 1, 2026, which bases publisher compensation on how AI search actually uses a page rather than a flat per-crawl fee.

Not every large site is taking the paywall option. Quora hardened the same day from a redirect to a flat 403 for all four probes — GPTBot, ClaudeBot, PerplexityBot and the browser alike — and its robots.txt, readable to the panel for the first time, disallows all three crawlers by name and bars AI-training use outright. Reddit moved separately, on its own infrastructure rather than Cloudflare's, adding a 403 to browser and PerplexityBot requests that had drawn a 200 the day before.

Three days later the panel caught a second large publisher rejecting the split outright, and one moving the other way on a single crawler. El País's homepage flipped from a 200 to a 403 for GPTBot, ClaudeBot and PerplexityBot alike between the September 13 and 16 captures. Stack Overflow moved just Anthropic's ClaudeBot — off Cloudflare's Pay Per Crawl paid track and onto a flat 403 block — while GPTBot and PerplexityBot stayed behind the same Cloudflare challenge as before: the panel's first capture of a crawler moving from paid access to blocked rather than the reverse.

This is now two capture days from the wire's own panel — September 13 and September 16 — not yet corroborated by another outlet's measurement.

tools · aeo

Why don't AI visibility tools agree with each other?

settled The record agrees. A change here would itself be news.

Because they measure different engines, query sets and windows. Two trackers reporting opposite trends for the same subject are usually both correct, and the disagreement is methodological rather than factual — which is why the fact of disagreement stays stable while every number under it moves.

Verified against the record 2026-09-04 · record last moved September 4, 2026

The clearest worked example the record holds is Reddit in August 2026. BrightEdge reported on August 16 that Reddit's AI citation surge had plateaued at a baseline several times higher than a year earlier, and that Google AI Overviews accounts for roughly 88% of Reddit's AI citation volume. Three days later a tracker reported an 86.4% collapse. Both were right.

Promptwatch's own per-engine breakdown resolved it: ChatGPT Search fell 86.4% (3.83% to 0.52%), while AI Overviews slipped 11.3% and AI Mode 30.5% over the same window. The surface carrying ~88% of the volume barely moved. A decline reported widely as Reddit disappearing from AI search was one engine changing how it assembles background queries.

A separate, single-source measurement of that query-writing step shows why the change moved so much: brands ChatGPT named in its own background search query reached the final answer 68.9% of the time, against 2.1% for brands surfacing only on retrieved pages — a roughly 33x gap — and only 3.1% of 3,554 retrieved pages earned a citation at all. Read against Promptwatch's finding, the two measurements describe the same mechanism nine days apart: the query-writing step, not the page, is what the Aug 8 change actually altered.

Engine-to-engine spread inside a single study is wide enough to swamp most between-study comparisons: Writesonic measured "ghost citation" rates from 52% (Perplexity) down to 19% (Microsoft Copilot) across seven engines and roughly 16 million brand appearances. A figure quoted without its engine is not a figure.

The model underneath moves fast enough that naming the current one dates the paragraph. AI Mode's underlying Flash model has swapped four times since July 2026 — 3.5 Flash-Lite, then 3.6 Flash, 3.7 Flash on August 14, and 3.8 Flash on September 2 — each arriving roughly three weeks apart, which is faster than most visibility studies are conducted and published.

A day after the 3.8 Flash swap, two SEO practitioners reported AI Mode responses running on it showing markedly fewer citations than before — screenshots of several top-of-funnel queries returned no source links at all. Google confirmed it the next day: Robby Stein, Search's VP of product, said on X that citations were not working as intended and a fix is coming. Still no query sample or scope from either side, but the mechanism is now acknowledged rather than anecdotal — a bug, not a redesign.

A same-day finding complicates measurement further, upstream of any tracker: ChatGPT doesn't answer from one index. Peec AI reports OpenAI has built its own retrieval system, internally named Labrador, with separate vertical indexes for web, PDF, video and news — but a since-removed `result_source` field visible in ChatGPT's server-side events between May 21 and July 21, 2026 showed answers still drawing on Google, Microsoft's Bing, and two third-party scraping services alongside it. A dashboard built to approximate ChatGPT exposure from Bing rankings alone is reading one of at least four backends, not the whole surface.

aeo · research · tools

Once I'm cited, do I stay cited?

settled The record agrees. A change here would itself be news.

Not reliably. For most prompts a small stable core of domains holds while everything around it rotates fast — the largest published measurement found Google replaces 56% of its AI-answer citations every week, ChatGPT as much as 74%. The instability is structural, not a sign anything is broken, and three separate measurements had already caught pieces of the same mechanism before this one supplied the base rate.

Verified against the record 2026-09-11 · record last moved September 11, 2026

SISTRIX re-sampled 82,619 prompts weekly for 17 weeks (1,548,213 snapshots, six countries, three surfaces) and found the headline churn sits on top of real structure: 86% of prompts keep a stable core of a few domains, with everything outside it rotating at 89% per week. On Google AI Overviews specifically, 53% of prompts saw no source change at all across the full window — aggregate churn and per-prompt permanence are both true, describing the tail and the core of the same answer.

The structure has consequences the record had not connected before. AI Overviews and AI Mode — one company, the same query, the same week — cite different domains 83% of the time, so a citation-share number measured on one Google surface is not a reading of the other. And permanence varies sharply by content type: only 1.4% of cited news articles held their spot across all 17 weeks, against 43% of brand queries that kept the brand's own domain present throughout even as the co-citations beside it rotated at 70% per week.

Three items on this wire measured pieces of this before the wire had a name for it. Kevin Indig's Search Signals analysis found AI Mode's publisher mentions falling nearly by half in thirty days while retailer mentions tripled, across 15 of 20 verticals — the engine's preference changing. Promptwatch traced Reddit's 86.4% four-day collapse in ChatGPT citations to a change in how the engine assembles its background queries — the retrieval method changing. And Profound shipped a feature called Citation Decay on August 13, 2026 that tracks week-over-week citation counts per URL — a vendor building the instrument this research argues is necessary. Three different mechanisms; SISTRIX supplies the base rate they were all sampling from.

A model swap can also break citation outright rather than just shift its rate. Two days after Gemini 3.8 Flash rolled out to Google AI Mode on September 2, SEO practitioners found some top-of-funnel queries returning no source links at all; Google's Search VP Robby Stein confirmed it was a citation bug on September 4 and said a fix was coming. It is the sharpest confirmed instance yet of the standing caveat that citation share moves when the model underneath does — this time to zero, briefly, on one model.

A follow-up BrightEdge measurement narrows the mechanism to a single vertical and a single engine, and answers a question the churn rates alone couldn't: what happens to a slot once it's vacated. Reddit's share of ChatGPT's Education-category citations fell from 17.2% to 0% over two weeks in August, and the domains that took its place — state.gov, asu.edu, apa.org, aacnnursing.org — still held those positions five weeks later. The slot didn't revert to variety; it was claimed by a narrower, more institutional set and stayed claimed. BrightEdge attributes the drop to Reddit's robots.txt policy; this wire's own August 27 reporting found a ChatGPT query-behavior change that preceded the decline by about six days, so the cause is still unsettled even as the outcome — persistence, not reversion — replicates.

aeo · research · tools

What can a publisher actually control about AI answers?

moving Actively changing — a rollout, a deadline, or a live fight.

Appearance and training are separate controls and no single lever covers both. Google now offers the only control that removes a site from AI Overviews and AI Mode without touching ordinary Search — it reached all sites worldwide on August 31, 2026, and it does not cover training.

Verified against the record 2026-08-31 · record last moved August 31, 2026 — nothing has moved it in 17 days

Four levers exist and each has a different scope. Choosing wrongly costs more than doing nothing: three of the four take ordinary Search results with them.

The price of the one that does what publishers asked for is stated plainly by Google — opt out and "you won't receive any traffic or impressions from these features" — against AI Overviews appearing on 43% of Google searches in July 2026, up from 15% a year earlier.

A regulator is already pushing on the gap. The UK CMA's Publisher Conduct Requirement, imposed June 3, 2026 and legally binding under the digital markets regime, requires Google to let publishers opt out of training, fine-tuning **and** the grounding behind AI Overviews and AI Mode, with nine months to comply. The Search Console control answers one of those three.

Google's own guidance has not caught up with its own control. "AI features and your website" — the Search Central page telling publishers how to approach inclusion — named `nosnippet` 18 times and the Search generative AI control zero times when checked on August 27, 2026. A publisher following the documentation rather than opening Search Console would not learn the control exists.

google-search · aeo · law-money

Does llms.txt do anything?

thin One measurement or one source, without corroboration.

For Google, no — its documentation says outright that no AI text file is needed to appear in its AI features. For every other major engine there is no published answer either way, and the one large-scale measurement found 97% of llms.txt files were never read by an AI crawler.

Verified against the record 2026-08-31 · record last moved August 31, 2026 — nothing has moved it in 16 days

The evidence for whether the file is read is one vendor study and one vendor's documentation. That is thin for a tactic this widely recommended, and the thinness is the finding: nobody has published a reason to make the file, and only one company has published a reason not to.

Ahrefs, across roughly 15 million data points, found 97% of sites' llms.txt files were never read by an AI crawler, and that adding schema markup produced no measurable lift in AI citations across 1,885 pages over 30 days. The same research found 88.46% of AI citations still traced to pages in the general search index — the tactic that works is the old one.

The spec itself is still moving. Jeremy Howard published a version 2 update on August 10, 2026, its first revision since the format launched in 2024, adding `rel="alternate" type="text/markdown"` and `rel="describedby"` link relations so agents can find a page's Markdown version.

Google's John Mueller added a behavioral data point rather than a measurement: on his own test sites, "the only crawlers who claim to accept markdown are SEO tools" — no major AI bot did. He recommends logging accept headers before building a markdown version at all, which is the same conclusion the Ahrefs read-rate finding points to by a different route.

Common Crawl's analysis of 584,107 llms.txt files in its July 2026 archive found 68% were produced by templates or SEO plugins — Wix alone accounts for 41% of the corpus — with only about half fully following the spec's structure. It doesn't measure whether crawlers read the files, but it corroborates the same direction from a different angle: adoption is running mostly on CMS-plugin autopilot rather than deliberate authoring.

aeo · tools · llmeo

Are AI answers costing me traffic?

contested Credible sources disagree, and the disagreement is unresolved.

It depends on what you publish, and part of the disagreement is instrumentation. Commerce reports AI referrals as growth; nonprofit and health publishers report the steepest declines; general-news outlets diverge — Le Monde reports none seven weeks after AI Overviews launched in France, while People Inc.'s Google search traffic fell 40% — and Google's own figure arrives without methodology. The one non-vendor academic estimate, a University of Washington study of Wikipedia, lands far below the self-reported figures at roughly 5%. Before any of it is compared, 22.4% of AI Overview traffic is misattributed away from Organic Search.

Verified against the record 2026-09-16 · record last moved September 16, 2026

The disagreement is real and it sorts by what a site sells. Shopify told its Q2 earnings call that AI-driven traffic and orders tripled year over year, with half of AI-referred sessions landing directly on product pages — 2.5 times the rate of traditional search. Over the same period Blood Cancer UK's leukaemia information page lost 53% of its page views. Both are first-party reports from parties with no incentive to overstate against their own interest.

One widely-cited framing does not survive the record. Zero-click searches hit a record high in June 2026 — 40% of U.S. searches sent a click to the open web — but the same report put Google AI Mode at 0.13% of U.S. search-related visits, usage it described as plateaued or dipping. A surface used by roughly one in a thousand visits cannot be the cause of a decline that broad, whatever else is.

Measurement sits underneath all of it. A nine-month GA4 study of 51,200 clicks found an average 22.4% of AI Overview traffic recorded as Direct rather than Organic Search, and Similarweb found 58.8% of ChatGPT referral traffic lands on publisher homepages even though 65% of cited URLs point two or three folders deep. A citation and the session it produces are not measured in the same place, which is why two honest parties can count the same channel and disagree. See the measurement-disagreement claim.

The response has a price tag now too. Similarweb's Top 100 Media publishers spent an estimated $113 million on Google paid search in July 2026, up 41% year over year and 274% over three years — Forbes alone spent $72.2 million, about two-thirds of the total, even as its own organic traffic fell 26.7% year over year. Buying back the clicks AI answers are keeping is a growing spend line, not an absorbed loss.

Le Monde's CEO complicates the uniform-decline framing directly: seven weeks after AI Overviews launched in France, the outlet logged 24,000 new subscribers in each of July and August 2026 and 7% growth in English-language traffic for H1 2026 versus H1 2025 — while French publishers' group APIG reported a 33-38% search-referral loss industry-wide. Louis Dreyfus's own account of the gap: search drives only about half of a major news site's visits, and the AI Overviews damage he's tracking is concentrated in service-content categories — health, food, automotive — down as much as 50%, not general news.

The only independent, non-vendor measurement in the record lands well under every self-reported figure above. A University of Washington paper estimates AI Overviews cut monthly search referrals to English Wikipedia by roughly 5% since the feature became Google's US default in 2024 — a figure the authors themselves revised down from an initial draft estimate near 15% after switching from daily pageviews to monthly referrals and adding international controls. Google disputes even the smaller number, saying Wikimedia's public referral data mixes in traffic from every search engine and can't isolate Google's effect; the authors counter that other engines carry too little share to account for the gap. The revision is itself informative: the same underlying effect produced a threefold-different headline depending on methodology, which is reason to treat every self-reported percentage above as provisional in the same way.

France now has independent, site-level evidence rather than one CEO's account. Ahrefs tracked 963 high-traffic French domains through Google Search Console for 28 days before and 9 days after AI Overviews launched in France on July 22, 2026: sites most exposed to AI Overviews (over 20% of their queries triggering one) lost 23.1% of CTR, against a 5.7% median CTR loss and 8.8% median click loss across all tracked domains, with nearly 48% of sites losing more than 10% of daily clicks within nine days. This is not in conflict with Le Monde's CEO — it corroborates his own account of where the damage sits. Ahrefs' worst-hit category was health (22.2% AI Overview rate, 20.4% CTR decline), a service-content vertical, matching the categories — health, food, automotive — Dreyfus named as carrying the losses he does see, while general news is where he reports none.

seo · google-search · research

What on-page work actually moves AI citation?

settled The record agrees. A change here would itself be news.

The levers that measure are the ones that were already SEO — being in the search index, ranking well, being reachable in plain HTML, and keeping pages current. The AI-specific file formats measure near zero.

Verified against the record 2026-09-09 · record last moved September 8, 2026

The single largest number in the record points back at classic search: 88.46% of AI citations traced to pages already in the general search index, and 76% of the passages Google reused 100 or more times came from pages already ranking first organically. Google's own position is the same — John Mueller said there is "nothing really special you need to do for generative AI responses in search".

That is not the same as saying nothing is specific to AI. Three levers measure, and all three are mechanical rather than semantic: whether a crawler can reach the page at all, how recently it changed, and whether the passage is shaped to be lifted. The AI crawlers are stricter than Googlebot on the first — GPTBot, ClaudeBot and Bingbot found none of a site's JavaScript-injected internal links across 41 days, and recovered 250 pages within 48 hours of the links being converted to HTML.

The tactics named after the category are the ones that do not measure. An analysis spanning roughly 15 million data points found 97% of llms.txt files were never read by an AI crawler, and adding schema markup produced no measurable citation lift across 1,885 pages over 30 days. See the llms-txt-effect claim, which holds that question's own open threads. A separate study re-tested the field's original 2023 citation-lever effect sizes directly against ten modern AI engine families and found none of them moved citations any longer; a content score built from those levers correlated with actual citations at just 0.11 within-query, positioning it as a quality filter rather than a citation predictor.

aeo · geo · seo · research

Can I buy my way into AI answers?

moving Actively changing — a rollout, a deadline, or a live fight.

You can buy a slot, but not a citation — they are separate games. Ads run on about a quarter of ChatGPT's commercial prompts and nearly one in three commercial AI Mode queries, and only 3.63% of the advertisers whose ads ran were also cited as a source in the answer above them. The ad stack is scaling faster than the reporting under it.

Verified against the record 2026-09-17 · record last moved September 17, 2026

The ad formats keep widening inside AI Mode, not just below it. Google confirmed testing regular Search-campaign text ads — for now limited to exact and phrase match keywords carrying "explicit and direct" purchase intent — inside AI Mode responses, on top of the shopping ads carousel that arrived August 31. Google has not said how long the test will run or how many advertisers are included.

The paid layer arrived quickly and is still moving. ChatGPT Ads launched in the US on February 9, 2026, reached nine markets by August 17 — its first Spanish- and Portuguese-language ones — and roughly 40 countries two days later when 31 European markets went live at once. The business crossed $1 billion in annualized revenue run rate in under 200 days after launch, per OpenAI, and self-service Ads Manager access opened to India, Europe, the Middle East and North Africa on August 31 — previously a managed-sales and partner-led market. Ads remain limited to Free and Go plans. On September 10, Amazon Ads began piloting a third route to that inventory: buying ChatGPT Ads placements through Amazon's own DSP for select US advertisers, with Amazon managing the buy and OpenAI still controlling how and where ads appear. No timeline was given for expanding beyond the pilot. The next day, Adform became the first European-headquartered name on OpenAI's technology platform partner roster, letting its clients manage ChatGPT campaigns alongside their other Adform-run media; Volkswagen and Vodafone were named as early clients. On September 16, OpenAI moved distribution into the software marketers already run their day in: HubSpot became ChatGPT Ads' first CRM integration and Shopify its first ecommerce integration (U.S. merchants first, other countries from September 23), and OpenAI began testing Sponsored Agents — a labeled ad format that opens a separate conversation with the advertiser's own AI agent instead of sending a click to its site — with select US advertisers. Ads Manager also gained per-platform targeting and reporting (Android, iOS, web, desktop) and 7/14/30-day attribution windows, up from a fixed one day.

The format itself is moving into the answer, not just below it. As of August 31, 2026, Google began showing a shopping ads carousel inline within AI Mode's response, alongside the existing separate ad unit at the bottom of results — reported from a single ad-format tracker, with Google neither confirming the change nor stating a rollout scope.

Paid presence and cited presence are close to unrelated, which is the finding that matters for anyone treating ads as an AEO shortcut. Across more than 50,000 commercial prompts in 20 niches, 3.63% of advertisers running an ad were also cited as a source in the answer above it, and 14.35% of the ads shown were unrelated to the surrounding conversation. The same separation shows on Google: an analysis of AI Mode text ads found advertisers rarely among the answer's cited sources.

Two coverage figures circulate and they are not in conflict — they use different denominators. SE Ranking found ads on 25.94% of commercial prompts; Adthena, measuring all US queries rather than commercial ones, found 4.47% at an average of 1.06 ad items per response, against a 3.53-item average on Google's AI surfaces. Read the first as ad load on the queries advertisers want and the second as ad load across everything.

Category concentration is shifting under retail's early lead. Sensor Tower's April-August window found financial services growing from one top-100 advertiser to four of the top ten spenders, and from 2% to 13% of US ad spend, while shopping and retail's share fell from 37% to 21% and travel rose from 5% to 11%. Comscore's new panel-based tracking corroborates the travel shift from a different angle: sponsored links on hotel-related ChatGPT prompts rose from 6% to 24% between March and May, measured from real opted-in user prompts rather than synthetic tests.

Reporting lags the levers throughout. Campaign-level platform targeting shipped in August into an Insights view that still groups results as Mobile and Desktop, so a web-only campaign reports entirely as Mobile. View-through conversions appear as their own column but are excluded from CPA, bidding and billing. Automated bidding became the default for new ad groups carrying no guarantee against a CPA, CPC or ROAS target. Third-party tooling is currently the only cross-platform view of what is actually running.

engines · tools · aeo · google-search

Will the courts stop AI engines from using my content?

moving Actively changing — a rollout, a deadline, or a live fight.

Not so far. The one concluded case produced a price rather than a prohibition — roughly $3,000 per work — while 143 AI copyright suits remain pending nationwide, 27 against OpenAI and 15 against Microsoft. Courts are reaching opposite results on near-identical theories, a First Amendment defense has appeared from three sets of defendants in six months, the Justice Department has now filed its first formal position on the fair-use question underlying nearly all of them — backing it — a new suit has added a trademark-dilution theory that stands apart from copyright altogether, and the first federal appeals ruling on an AI-output copyright question, from the Ninth Circuit, went to the AI company.

Verified against the record 2026-09-17 · record last moved September 16, 2026

One case has produced a number. The Bartz v. Anthropic book-piracy class settlement became effective August 20, 2026. Class counsel now says the first payout — $2,203.56 per work — goes out on or before Nov. 15, 2026, later than the Sept. 17 date estimated when the settlement took effect: consolidated payout statements go out Sept. 4, then a 30-day window for class members to contest them runs before money moves. A second payment, funded by a further $450 million Anthropic owes plus interest, is expected to bring the total toward the roughly $3,000 per work estimated earlier. Two appeals filed since challenge only attorneys' fee awards and by the agreement's terms do not affect the payout. It is the first figure in the record that prices training on pirated books, and it arrived as a settlement rather than a ruling — so it binds nobody else.

The theories are splitting rather than converging, sometimes within days. A federal judge dismissed most of Google's DMCA anti-circumvention claims against SerpApi on the reasoning that Google's anti-bot system protects ad revenue rather than a copyrighted work; days later a different judge let Reddit's near-identical DMCA claims against Perplexity and SerpApi proceed. Contributory infringement has moved the other way and closed: the New York Times, Daily News and Ziff Davis had those claims dismissed with prejudice against Microsoft after the Supreme Court's Cox Communications decision foreclosed the theory.

Agentic browsing got its first appellate answer, and it favored the agent. The Ninth Circuit reversed an injunction barring Perplexity's Comet browser from Amazon, finding Comet unlikely to violate the Computer Fraud and Abuse Act because it acts on the user's direction rather than accessing servers on its own.

A defense posture is forming in parallel. Musk, Tesla and Warner Bros. Discovery raised a First Amendment defense in an AI-copyright dispute in March 2026; Anthropic added one in the Gilbert case on August 20; Perplexity raised one in Reddit's DMCA suit on August 28. Separately, the pressure that has moved fastest is not copyright but competition — Judge Mehta, who already found Google's search business an illegal monopoly, said in a hearing that Google's use of publisher content in AI Overviews "seems really unfair".

The federal government took a side for the first time on September 1. The Justice Department filed a Statement of Interest in the OpenAI copyright litigation arguing that training on copyrighted text is a transformative fair use, and calling plaintiffs' "market dilution" theory — that AI outputs harm a work's market merely by resembling its genre — "deeply flawed." A Statement of Interest under 28 U.S.C. § 517 states the government's view; it does not bind the court deciding the case.

A new theory arrived September 4, outside copyright entirely. The Seattle Times and Newsday's joint suit against OpenAI and Microsoft adds a Lanham Act trademark-dilution count, arguing that ChatGPT and Copilot outputs which hallucinate content and misattribute it to the papers tarnish their marks — a claim that would survive independently of however the fair-use question is decided.

The DOJ's position didn't stay in the case it was filed in. On September 4, ROSS Intelligence — a defendant-appellant in Thomson Reuters v. ROSS Intelligence, one of the earliest AI-training fair-use rulings, now on appeal to the Third Circuit — filed a letter of new authority citing the DOJ's Statement of Interest and arguing it applies broadly to AI training, not only to generative AI. It is the first sign of the DOJ's position being invoked outside the case it was filed in — by a party, not yet by a court.

A new track opened September 9, aimed at boards rather than the underlying infringement. Katelyn Gray filed a shareholder derivative suit against Satya Nadella, other Microsoft directors and officers, and Microsoft itself, arguing they approved or were exposed to copyright infringement risk and misrepresented it to shareholders — the third such "copyright shareholder derivative" suit against Microsoft's board, expected to be consolidated with the two before it into one stockholder derivative action. The theory tests board oversight of copyright risk rather than whether the underlying training was infringing.

A dispute over whether a promise not to sue can pre-empt a fair-use ruling opened September 11. Anthropic opposed Daniel Gilbert's motion to dismiss the fair-use counterclaim it filed against him in his pro se suit over *Hacking World of Warcraft*, arguing his covenant not to sue over "training copies" does not moot the counterclaim because it rests on his still-pending infringement claim and his position that the court cannot weigh Anthropic's purpose for copying — not on fear of a second suit.

A second appellate answer arrived September 16, again favoring the AI company but for a narrower reason than the lower court gave. The Ninth Circuit affirmed dismissal of a claim that GitHub Copilot violated copyright law by stripping copyright management information (CMI) — the ownership and licensing tags attached to code — when it generated new code, agreeing that producing a new AI output isn't the same as removing tags from an existing copy. But it rejected the district court's own reasoning for getting there, calling its "identicality requirement" a misnomer and declining to endorse it: the first federal circuit-court ruling on this DMCA question for AI outputs.

The litigation calendar moved in two directions at once on September 15-16, both in cases against Anthropic. Summary judgment briefing closed in Concord Music v. Anthropic I, with a hearing set for October 21, 2026 before Judge Lee — one of only three AI copyright cases nationwide, alongside In re Mosaic LLM Litigation and In re OpenAI Copyright Infringement Litigation, to reach that stage this year, and the first of the three scheduled to be heard. Separately, Judge Pitts set a scheduling order for Cambronne v. Anthropic — the case brought by authors and publishers who opted out of the Bartz book-piracy settlement — putting non-expert summary judgment motions due May 13, 2027 and trial not until 2028, years behind the Bartz class members now waiting on their first settlement payout.

A second newspaper publisher sued OpenAI and Microsoft in as many weeks. Times Publishing Company, the corporate publisher of the Tampa Bay Times, filed suit September 16 in the Southern District of New York, alleging copyright infringement, vicarious copyright infringement and DMCA copyright-management-information removal — the same theories the Seattle Times and Newsday raised jointly on September 4. Per the complaint's own count, it is the 27th copyright suit against OpenAI and the 15th against Microsoft, out of 143 filed industry-wide, up from the 24-against-OpenAI, 137-total count in the record three weeks earlier. The case is expected to be assigned to Judge Stein and folded into the consolidated In re OpenAI Copyright Infringement Litigation MDL, then stayed pending the summary judgment briefing already underway there.

law-money

I'm getting cited. Am I getting recommended?

settled The record agrees. A change here would itself be news.

Not necessarily — being cited, being named and being recommended are three different outcomes with different drivers, and the record separates them. Across 1,094 ChatGPT categories the most-cited domain was the most-mentioned brand only 20.8% of the time. Citation measures whose page an engine used; recommendation tracks how well the engine already knows your brand.

Verified against the record 2026-08-25 · record last moved August 24, 2026 — nothing has moved it in 23 days

The three outcomes come apart at every step. An engine can lift a passage from your page and name a competitor in the sentence it supports; it can name you without linking you; it can recommend you having never cited you at all. A dashboard reporting one number cannot tell you which of the three you bought.

What moves recommendation is mostly not on-page work. Models searched for brands they already knew 3.2 times more often than unfamiliar ones — 55.7% versus 17.4% across 66 buyer prompts — and 63% of brand-specific searches surfaced one of each model's five most-familiar brands. Where competing products were otherwise identical, the well-known brand was recommended 100% of the time.

Two findings look like they disagree about reviews and do not. A market census of 4,776 venues found star rating had no effect on whether a venue appeared at all; a controlled study found a rival needed less than a 0.1-star edge to overturn an incumbent's recommendation. Those are different stages of the same funnel — rating does not get you into the candidate set, and decides between candidates once you are in it. Having a business website nearly doubled the odds of the first.

The uncomfortable finding for this discipline sits in the same study: when every brand adopted the same marketing tactics, the incumbent's advantage collapsed from a payoff of +0.802 to +0.007. Tactics that everyone runs stop being an edge and become the price of entry — while brands that opted out received no recommendations at all.

aeo · research · engines