00The week in one paragraph
This was the week the industry’s two underlying questions came due at once: who gets paid when machines read the web, and what the open market is actually worth. Publishers spent it arguing over crawlers, markdown and agentic curation — People Inc. disclosed Google search traffic down 40% YoY and still won’t pull the crawler trigger, Cloudflare put bots at over half of all web traffic, and PubMatic and Optable wired their sell-side and buy-side agents together so publisher first-party data can answer a campaign brief in real time — while Google quietly shipped Buyer Direct into the ad server most large publishers already run, at a fee publisher conversations put near 10%. Beneath that, ad tech repriced itself: Nielsen is taking DoubleVerify private for $2.15bn, leaving both major independent verification firms in private hands; The Trade Desk grew 3% and lost over 20% of its value after hours; AppLovin grew revenue over 50% and fell over 20% anyway; and Magnite’s CTV line grew 36% as decisioning moved toward the sell side. OpenAI bolted product carousels onto ChatGPT ads against a reported $2.5bn ad revenue target, New Jersey’s data broker law went live with $50,000-per-record penalties that reach publishers selling their own first-party data, and the IAB said the quiet part about AI visibility scores: below 50 queries, the data characterises nothing. The thread: the machines reading and buying the web are arriving faster than the rules — or the prices — for what they take.
01Google Buyer Direct lands in the ad server you already run
Google rolled out Buyer Direct quietly, and it deserves more noise: it lets AI agents automate direct sales by reading availability, live bookings and pricing straight out of the ad server — the ad server most large publishers already run (AdExchanger, 7 August). Publisher conversations put the fee near 10% — cheaper than the programmatic chain. Ad tech veteran Gareth Glaser on why it works: “no middlemen, no hops, there’s massively reduced latency. There’s tremendous transparency because the ad server really knows what’s going on.” The opportunity is less friction on direct deals — more working media without rebuilding anything. The threat is structural: Google could only build this because it already owns your ad server, and as Glaser puts it, now “programmatic needs to earn what it’s doing.” The target isn’t your direct business; it’s everything you currently pay a supply chain to do.
Publisher takeaway
An agent transacting against your ad server is only as good as what your ad server knows, so the preparation is unglamorous: clean up GAM before Buyer Direct reads it. That means a real product catalogue — line-item types, targeting presets and creative specs that describe what you actually sell — and rate cards that reflect prices you would defend, because an agent reading availability and pricing will quote whatever it finds, including the stale placeholder rates nobody has touched since 2023. Audit forecast accuracy too; automated direct deals against inventory you systematically over- or under-forecast is how you end up with makegoods negotiated by software. Treat the 10% fee as a claim — it comes from publisher conversations, not a published rate card — and model it against your true effective programmatic take, hops included, before deciding which inventory to expose. Then weigh the dependency honestly: this is a genuinely cheaper pipe, offered by the company that already runs your ad server, and every deal that moves through it deepens that position. Route incremental and mid-tier direct demand through it first; keep your largest direct relationships human until you’ve seen a quarter of reporting.
02People Inc. won’t block Google — because its hedge is working
People Inc. is down 40% YoY in Google search traffic and 22% in unique visitors, yet CEO Neil Vogel won’t pull the crawler trigger: “If we were to turn off AI, we would turn off search… we’re not there yet” (Digiday, 4 August). The more useful number is the hedge: non-session revenue grew 16% YoY, from $108M to $125M, and now makes up 43% of digital revenue, up from 39% — Apple News licensing, AI deals, events, and the company’s D/Cipher targeting tool. Session revenue held flat while sessions fell, because rates rose. Blocking crawlers is a negotiating chip People Inc. can’t yet afford to spend — so it is building the position from which the question stops mattering.
Publisher takeaway
This is the clearest public model yet for surviving the referral collapse: don’t win the crawler fight — stop needing to. Copy the accounting before you copy the strategy. Split your revenue into session-dependent and non-session lines and put the ratio in front of your board quarterly; most publishers cannot currently say what their 43% number is, and you can’t manage a hedge you don’t measure. Then notice the quieter fact in People Inc.’s numbers: session revenue held flat on falling sessions because rates rose. That is yield work — floor discipline, curation, direct and PMP mix — offsetting a traffic decline, and it’s available to you regardless of whether you ever sign a licensing deal. Set a target of raising revenue per session faster than sessions fall, and treat that as the operating metric for the next four quarters. On the crawler question itself, take Vogel’s framing literally: the block is a chip, and a chip is only worth something if you could actually play it. Build per-crawler control now — selective, staged, reversible — so that when your non-session share makes blocking survivable, the switch already exists. Deciding to block and becoming able to block are separate projects; start the second one first.
03Stealth crawlers are routing around your robots.txt
Stealth crawlers mask their identity, ignore robots.txt and mimic human browsing patterns via residential IPs — which means your blocklist and your licensing deals are both being routed around (Digiday, 4 August). The scale: Cloudflare now puts bots at over half of all web traffic, AI scraper traffic grew 597% across 2025, scraping hits nearly 20% of site traffic at the median organisation, and People Inc. alone blocks more than 30,000 user agents a day. News/Media Alliance CEO Danielle Coffey frames the fix: “Bad actors, bad bots must identify themselves, and then when they do, we can stop them.” New York’s Stealth Crawler Prohibition Act attaches up to $15,000 a day per violation. Every robots.txt-based monetisation strategy assumes bots identify themselves; a growing share simply don’t.
Publisher takeaway
Robots.txt is a request, not a control, and every strategy built on it — blocking, pay-per-crawl, licensing tiers — inherits that weakness. Three moves follow. First, reconcile intent against reality: pull server logs by user agent and ASN and compare them to what you believe you block. Most publishers who do this find crawlers they “blocked” months ago still taking content, and that’s a fixable engineering problem, not a policy debate. Second, invert the model — the piece’s practical conclusion is that latency, whitelisting and monitoring beat blocklists. A default-deny posture with an explicit allowlist for the crawlers you’ve chosen (search, verified partners, licensed AI) fails safe; a blocklist fails open every time a new stealth agent appears, which is daily. Rate-limiting and tarpitting suspect traffic costs the scraper compute without requiring you to prove identity first. Third, put a number on it: bot management is now a real line item — People Inc. is fighting 30,000 user agents a day — so cost it, budget it, and count it against whatever any licensing negotiation offers you. And keep your reported audience clean: if half of web traffic is automated, buyers will start asking how much of yours is. Have the human-only number ready before they do.
04New Jersey’s data broker law is already live — and it covers your first-party data
Signed 30 June 2026 and already in force, New Jersey’s law bans the sale of sensitive data — health, precise geolocation, financial, biometric, immigration status, children’s data — and, critically, invents a “data collector” category that captures first-party data sellers, including media companies and retailers with loyalty programmes (AdExchanger, 3 August). Penalties run $50,000 per record with no cap; registration is due April 2027 with annual fees from $5,000 to $1.5M depending on how many New Jersey residents you market to. RTB House’s Charlie Simon: “It’s a multiplier with no cap. You don’t have to work the math far on a segment containing New Jersey residents before you’re past any revenue that segment ever produced.” Controllers may also carry liability for partners’ non-compliance, and litigation is expected.
Publisher takeaway
If you monetise first-party audience data, you cannot assume your direct relationship with the reader exempts you — the “data collector” category was written to reach exactly that assumption. Do the exposure maths Simon is pointing at: at $50,000 per record, uncapped, a 10,000-user segment containing New Jersey residents carries theoretical exposure of half a billion dollars, against segment revenue measured in hundreds. That asymmetry means this is triaged by segment, now, not scheduled for a 2027 compliance sprint. Inventory every audience product you sell — through your DMP, your curation deals, your SSP audience extensions — and answer two questions per segment: could it contain New Jersey residents, and does it touch any sensitive category, with health and precise geolocation the ones ad taxonomies wander into most casually. Kill or geo-carve anything that fails both. Then reread your partner contracts, because controller liability for partners’ non-compliance means a downstream buyer’s sloppiness can become your penalty; you want representations and indemnities that reflect that. Put April 2027 registration on the calendar with the fee schedule attached. And treat this as a template, not an outlier — the states copy each other, and the publishers who build segment-level geo controls once will not be rebuilding them per statute. Not legal advice; brief counsel this quarter.
05Buyer agents are finally making publisher first-party data sellable
PubMatic and Optable have connected their sell-side and buy-side agents end to end: Optable packages publisher first-party data into audiences, and PubMatic’s buyer agent reads a campaign brief and queries that data in real time (AdExchanger, 3 August). Roughly half of the ~30 agentic programmatic campaigns run on PubMatic so far came through this workflow, with Mediavine — 18,000+ publishers — and Livewire testing. Mediavine’s Charlie Morris put it plainly: publishers can “unlock a data asset that’s been there but untapped for quite some time.” This is not ad ops replacement; it’s ad ops leverage — curation that took a week of emails becomes a query.
Publisher takeaway
Keep the denominator in view — ~30 campaigns platform-wide is a pilot, not a market — but the direction matters, because it relocates the bottleneck. When curation was a week of emails, your constraint was sales capacity; when a buyer agent queries your data mid-brief, your constraint is whether the data can answer the question. Audit yours against that standard: are your segments described in taxonomy a machine can parse (IAB Audience Taxonomy, consistent naming, documented recency and provenance), or in labels like “engaged foodies Q3” that meant something to whoever built them? An agent skips what it can’t interpret, and skipped is invisible — you’ll never see the query you failed. The work is a data librarian’s quarter, not a platform migration: normalise segment names, attach definitions, record freshness, and make sure consent status travels with every segment, because an agent will happily query data you had no right to sell (see the New Jersey item above for what that now costs). Then ask your SSPs the direct question — what agentic buying workflows can read our audiences today, and what format do they want them in? If the answer is “none yet,” the taxonomy work still pays: it’s the same structure human curation buyers are already asking for.
06ChatGPT ads get real plumbing — and product carousels aimed at Q4
OpenAI now has CPC bidding, pixel tracking, conversion APIs and low spend minimums — the plumbing of a real ad platform. What it doesn’t have is attribution anyone trusts: AI Digital CSO Mary Gabrielyan says flatly, “We don’t have that answer for ChatGPT ads yet” (AdExchanger, 7 August). The same week, OpenAI added product carousels — multiple products in one placement, pulled straight from retailer feeds, with OpenAI’s systems deciding which products appear and whether to run single or carousel, currently one retailer per carousel (Digiday, 6 August). No ad units to build. The timing is not subtle: one agency exec says “OpenAI wants to make sure that these products are discoverable in ChatGPT conversations” ahead of the holiday season, and Adthena’s Ashley Fletcher notes Q4 “will be key” as advertisers set 2027 budgets. Reported ambition: $2.5bn ad revenue in 2026, $100bn+ by 2030. ChatGPT ads behave like intent-driven search, not display or CTV — and the carousel is a Google Shopping run-at built with no publisher layer in it.
Publisher takeaway
Two different exposures here, so separate them. The first is budget-side: because ChatGPT ads behave like intent search, they compete for the budgets attached to the queries that used to send you referral traffic — the money follows the intent, and the intent is moving. You can’t stop that, but you can measure your exposure: tag and segment ChatGPT-referred sessions now, however small, so you have a baseline before Q4 rather than an anecdote after it. The second exposure is sharper and belongs to commerce publishers: the carousel pulls straight from retailer feeds, which means the discovery step your reviews, gift guides and “best X for Y” pages monetise via affiliate links is being rebuilt without a publisher layer in it. If commerce content is a meaningful revenue line, stress-test it against a Q4 where some fraction of product discovery never touches the open web, and push your affiliate networks and retailer partners on whether publisher content will feed these systems and on what terms. Keep the claims flagged as claims — the $2.5bn and $100bn figures are reported ambitions, not results, and Gabrielyan’s attribution gap is real. But measurement gaps get closed. Plan for the platform that closes it, not the one that exists today.
07Nielsen takes DoubleVerify private for $2.15bn — verification loses its independence
Nielsen is acquiring DoubleVerify for ~$2.15 billion all-cash — $13.60 a share, a 30% premium, closing by Q1 2027, with DV keeping its brand and going private; combined revenue is projected above $4bn (AdExchanger, 6 August). With IAS already taken private by Novacap for $1.9bn in 2025, both major independent verification firms are now privately held. Nielsen CEO Karthik Rao pitches “a truly independent, end-to-end partner that connects trusted audience intelligence with verified media delivery” — but Nielsen operates inside the media supply chain it would now also verify, which complicates the independence claim underpinning the signals that grade your inventory. Why it happened is in the numbers: DV’s Q2 revenue grew just 3% to $193.8M, with programmatic activation down 1% — essential infrastructure, difficult public company (also covered by Digiday, 7 August).
Publisher takeaway
The signals that grade your inventory — viewability, brand safety, IVT — now have an owner with its own position in the chain, and no public-market disclosure obligations. Nothing changes on day one; closing is Q1 2027. What changes is your posture. First, stop treating verification verdicts as weather: keep your own parallel logs — ad server delivery, MRC-consistent viewability from a second source where you can afford it, IVT from your own traffic analysis — because when a DV classification costs you a deal, “we dispute it” only works if you have data of your own. Second, put a verification-dispute clause in your larger direct and PMP agreements now: which vendor’s numbers govern, what the reconciliation process is, and what happens when vendor and publisher numbers diverge past a threshold. That language is easy to get while everyone assumes the vendors are neutral referees; it gets harder once someone’s number costs someone money. Third, watch pricing — private ownership plus a shrinking independent field is the classic setup for fee increases, and publishers increasingly carry verification costs directly through wrapper and SSP integrations. When your SSP’s verification line item moves in 2027, you’ll want to have benchmarked it in 2026.
08The Trade Desk grows 3% — and the mix is a demand map
The Trade Desk reported $715M in Q2 revenue, up just 3% YoY, and the stock fell over 20% after hours (AdExchanger, 7 August). CEO Jeff Green didn’t dress it up: “Our revenue growth is below our expectations and below the standard we hold ourselves to,” pointing at macro pressure on big advertisers — auto and CPG especially. The mix tells publishers more than the headline: brands outside the Fortune 500 grew 50% YoY, EMEA and APAC grew over 30%, and audio was the fastest-growing media type at 7% of Q2 spend. On the 20% take rate that has held for a decade, Green was firm: “we’re extremely confident that we’re adding more value than we cost.”
Publisher takeaway
Read the mix, not the headline, because the mix is a demand map for your own pipeline. The softness is concentrated in Fortune 500 budgets — auto and CPG — which is exactly the demand most premium publishers’ direct and PMP pipelines are built around. The growth is mid-market brands (+50%) and EMEA/APAC (+30%), two demand pools most publisher sales operations are not shaped to serve: mid-market buyers won’t sit through a six-week RFP for a $40K test, and international demand never meets your sellers at all. So adjust the machinery. Build a lighter path to yes for mid-market money — packaged PMPs, self-serve or near-self-serve deal creation, standard products a smaller buyer can activate in days — and make sure your inventory is well-represented in the curated marketplaces and deal libraries where those buyers actually shop, because they buy off the shelf, not off the pitch deck. If you have meaningful non-US audience, stop treating it as remnant: geo-split your reporting and price EMEA/APAC deliberately rather than letting open-market floors set it. And note audio at 7% and climbing — if you have podcast or streaming-audio inventory that isn’t programmatically accessible, that’s the fastest-growing line on the largest independent DSP going unaddressed.
09Magnite doesn’t want to be a DSP — it wants the decisioning layer
“More valuable decisions are moving toward the supply side” — that’s the thesis of Magnite’s quarter. CEO Michael Barrett is pushing the SSP into decisioning work that historically sat with DSPs — packaging inventory, audience enablement, optimisation — via SpringServe, while insisting “we, in no way, shape or form are trying to replace the DSP” (AdExchanger, 5 August). The numbers behind it: Q2 revenue $193M, up 11% YoY; contribution ex-TAC $190M, up 17%; CTV ex-TAC $97M, up 36%. For publishers this is the good version of consolidation: more decisioning value moves to the sell side without locking you into one buyer.
Publisher takeaway
When decisioning moves sell-side, the economics of it move toward whoever holds the seat — and the question is whether that’s you or your SSP. So take the thesis seriously and interrogate your own SSP relationships against it. Ask each partner three things in your next QBR: what packaging and curation capabilities can we drive ourselves rather than have done to us; what audience enablement do you support with our first-party data, and who owns the resulting segment; and what does your decisioning cost — because “value moving to the supply side” has historically meant fees moving there too, and a curation or optimisation layer you didn’t ask for can quietly widen the spread between what buyers pay and what you clear. Get the fee stack in writing per deal type. The genuinely good news in Magnite’s framing is the non-exclusivity: sell-side decisioning that works across many DSPs strengthens your position in a way single-buyer lock-in never does — that’s the version to encourage. And the 36% CTV growth is the tell about where this contest is being fought first; if you have CTV or streaming inventory, the packaging and audience conversations are worth having now, while SSPs are competing to prove their decisioning layer on someone’s supply. Better it’s proved on yours, on terms you set.
10CTV’s real divide isn’t PMP vs. open market — it’s quality control
PMPs now carry 99.2% of programmatic CTV spend — and that was supposed to solve quality. It didn’t. The argument from Basis’ Ayse Pamuk: the PMP-vs-open-market debate is a distraction, because many deals carrying premium labels blend that inventory with lower-quality supply to hit volume commitments (AdExchanger, 5 August). The evidence: without safeguards, over 25% of CTV impressions may fail minimum quality standards, and the ANA found top-performing advertisers converted 54% of spend into qualified impressions versus 32.1% for the rest — despite similar PMP usage.
Publisher takeaway
If you hold genuinely premium CTV supply, this is a pricing problem disguised as a measurement problem: when premium and mid-tier inventory blur together in the same deal ID, the blend is priced as a blend, and you are subsidising someone else’s inventory. Transparent classification is therefore a revenue argument, not a hygiene one. Structure your deal IDs so that tiers never share one: separate deals by content tier, app bundle transparency and delivery environment, and document what each contains — series-level or genre-level content signals, verified app bundles, SSAI declaration — so a buyer can audit the claim. Refuse the volume-commitment trap that creates blending in the first place: when a deal’s volume ask exceeds what your genuinely premium supply can fill, the honest answers are a smaller deal or a second, clearly-labelled tier — not quiet dilution, which is how the whole market got here. Then sell with the ANA numbers in hand: a buyer converting 54% of spend to qualified impressions versus 32.1% is getting nearly 70% more delivered value from the same budget, which is the argument for your clean, verifiable deal carrying a premium CPM. The 25%-failure figure is a vendor-sphere estimate — use it to frame the problem, not as an audited fact. Your own logs are the audited fact; make them the product.
11Markdown for AI bots: cheaper to scrape, unproven to cite
Markdown versions of your site cut tokens processed by roughly 90% — stripping JavaScript, visuals and ads for cleaner, cheaper machine reading — and Time has gone furthest, serving ads to AI agents inside its markdown files (Digiday, 6 August). The case against: Promptwatch’s analysis of 1.6 million citations found markdown “did not increase” citation likelihood, Google’s own GEO guidance says it isn’t necessary, and ads in markdown haven’t improved advertiser results yet. The trade is concrete on one side and speculative on the other: you make yourself dramatically cheaper to scrape, in exchange for a citation lift nobody has measured.
Publisher takeaway
Worth testing; not worth committing to — and the test needs a design, because the failure mode here is shipping markdown site-wide on a hunch and discovering you’ve subsidised your scrapers. Note who captures each side of the trade: the 90% token saving accrues to the AI company’s compute bill, not to you, while the supposed benefit — more citations, more visibility — is exactly what Promptwatch’s 1.6M-citation analysis failed to find. So if you run the experiment, run it like one: a defined content subset, a citation baseline measured before launch (using at least the IAB’s 50-query minimum — see the visibility item below — so the data means something), and a review date at which the markdown comes down if the lift isn’t there. Gate the files while you’re at it: if you serve markdown, serve it to identified, allowed crawlers, not as a free public endpoint any stealth scraper can enjoy — otherwise you’ve built a discount aisle for the bots you’re elsewhere paying to block. Time’s agent-facing ads are the genuinely interesting variant, but even Time hasn’t shown advertiser results yet, so treat that as R&D, not a revenue line. If an AI company wants your content 90% cheaper to process, that efficiency has a price. Quote it in the licensing conversation; don’t give it away as an optimisation.
12The IAB’s AI visibility guidance — explicitly not a standard
The IAB’s “Measuring Visibility in the AI Era” proposes a 4Ps hierarchy — Presence (how often you’re cited), Prominence (where, and whether you’re highlighted or buried), Portrayal (sentiment and factual accuracy, including hallucinations about you) and Persuasion (post-citation click-through) — and is explicitly not a standard (AdExchanger, 3 August). IAB VP of AI Caroline Gigerenich says AI search still lacks the consistency to standardise. The practical guidance is the valuable part: a minimum of 50 queries before data characterises anything — below that it’s exploratory — and a hard line between “directional” and “decision-grade” measurement.
Publisher takeaway
This document is procurement ammunition, and you should use it that way this month, because the AI-visibility vendor market is selling hard into publisher anxiety right now. Every pitch that lands in your inbox claiming to score your brand’s AI presence now gets three questions with an IAB citation attached: how many queries underpin this score — anything under 50 is exploratory by the IAB’s own line, and a score built on a dozen prompts is an anecdote with a dashboard; is this directional or decision-grade, and will you put that word in the contract; and which of the 4Ps does it actually measure — most tools today count Presence only, while Portrayal (is the AI describing you accurately, or hallucinating?) and Persuasion (does a citation ever become a visit?) are where the commercial stakes actually sit. Internally, adopt the vocabulary even though the framework isn’t a standard: when your audience team reports “AI visibility,” make them say which P, on how many queries, refreshed how often — it will keep exploratory numbers out of board decks, where they otherwise calcify into targets. And note what the IAB is telling you by declining to standardise: the platforms’ behaviour is still too inconsistent to measure reliably, which means it is far too inconsistent to build strategy on. Measure monthly; commit to nothing.
13AppLovin grows 50%, falls 20% — a lesson in how slowly budgets move
AppLovin posted $1.9bn in Q2 revenue and roughly $1.3bn in net income, both up over 50% YoY — and the stock still fell over 20% after hours, because the expansion into ecommerce and consumer advertising is slower than investors wanted (AdExchanger, 5 August). CEO Adam Foroughi’s explanation is the useful part: “We’re deemed a new bucket, so a testing category. And to graduate up takes time.” Mid-tier ecommerce and consumer brands plan budgets one to four quarters out, search and social still hold most of the money, and Foroughi is deliberately prioritising “client acquisition rather than going for the most widescale possible adoption.”
Publisher takeaway
File this under expectation-setting, because Foroughi’s “testing category” line describes your new revenue products too. If AppLovin — with performance numbers most sellers would kill for and a growth machine behind them — still needs multiple quarters to graduate from a brand’s test bucket into its planning bucket, your new commerce media offering, your curated audience product or your newly programmatic newsletter inventory will not move faster. The mechanism is the same everywhere: non-endemic and mid-tier advertisers plan one to four quarters ahead, incumbent channels (search and social) hold the default budget, and a new line has to survive at least one full planning cycle as a test before it earns a recurring allocation. So build your forecasts on that physics. When you launch a new inventory type or data product, model two to four quarters of small test budgets before meaningful revenue, and say so upfront to your CFO — the most common way publisher innovation dies is not advertiser rejection but internal expectation mismatch, where a product on a perfectly normal adoption curve gets killed at month five for “underperforming.” Copy the sequencing too: land a handful of reference clients, get their results documented, and let those case studies do the scaling. Patience, in this market, is an operating discipline with a timetable — not a virtue.