All issues
Discovery Digest · 17 July 2026
Issue 09. Google's data moat cracks, AI reads what you never said, and the SERP becomes the store
TL;DR
The European Commission has forced Google to share anonymised Search data with rivals and AI chatbots, the first real crack in a 90% moat. A new study proves LLMs can infer your age, job and location from text that mentions none of it, and people block it only 28% of the time. Meanwhile Bing turns its Shopping results into the storefront, OpenAI ships an agent you delegate to, and product schema quietly gets cheaper visibility. Here is what changed and what to do about it.
01 · Search / Regulation
1. The EU forces Google to share its Search data with rivals and AI chatbots
- What
- On 16 July 2026 the European Commission adopted a final decision compelling Google to share anonymised Search data with rival search engines and, crucially, with AI chatbots that offer search functions. It is enforcement of Article 6(11) of the Digital Markets Act, aimed squarely at the query-data moat behind Google's 90%-plus European share. The Commission opened specification proceedings on 27 January 2026 after finding Google's own proposal stripped 90 to 100% of unique queries and excluded AI chatbots entirely. Eligibility is limited to search engines with verified investment plans that pass an independent audit, and permitted uses are tightly drawn: query understanding, ranking and retrieval (including grounding for AI answers), and indexing. Beneficiaries may not train general-purpose AI models on the data, use it for profiling or advertising, or systematically replicate Google's results. Google must anonymise in stages, starting with stripping direct identifiers such as usernames and IP addresses.
- When
- 27 January 2026 proceedings opened; final decision adopted 16 July 2026. First independent compliance audit within six months of sharing, annual audits thereafter, and a biennial review of the measures.
- How it shifts discovery
- This hands the raw ingredient of good search to challengers and AI answer engines without licensing a copy of Google's results, which over time could improve the query understanding and grounding behind competitor and AI-search surfaces. If you do GEO, treat this as a signal that non-Google answer engines are about to get materially better in Europe. Start tracking referral and citation share from AI search tools now, so you have a baseline before the data-sharing effects land.
- Questions to ask
- Which non-Google search and AI surfaces send us traffic or citations today, and do we have a baseline to measure change against?
- If rival engines get better query understanding, does our current Google-only optimisation leave us exposed?
- Are our EU discovery metrics broken out so we can spot a shift in the market?
- Sources
02 · AI Search / Privacy
2. New study shows LLMs infer what you never told them, and you can barely stop it
- What
- A study of 240 US participants confirms that models like ChatGPT can accurately infer age, location, occupation and relationship status from text that mentions none of them. When asked to rewrite the text to block the inference while keeping meaning, people succeeded only 28% of the time. Participants performed only a little better than chance at spotting what a model could deduce, and concern arrived only after the true inference was revealed. Benchmarked against the rewriters, ChatGPT itself blocked inference best, human rewrites landed 28% of the time, and the dedicated sanitisation tool Rescriber did worse than untrained humans. The strategy analysis is the practical part: paraphrasing was the most common approach and the least effective, while abstraction and added ambiguity worked better because they remove the underlying signal rather than reshuffling words.
- When
- Presented in the ACM Digital Library (paper reference 3772318.3791762).
- How it shifts discovery
- This moves the privacy conversation past data leaks to inference, which most marketing teams have not priced in. If your customer data, chat logs or personalisation prompts flow through an LLM, the model can deduce attributes you never collected, which is exactly the kind of risk a regulator notices. Audit what you feed models, prefer abstraction over paraphrasing when sanitising, and do not assume off-the-shelf privacy tools will save you.
- Questions to ask
- What personal attributes could an LLM infer from the customer text we pass through it?
- Are we sanitising by abstraction or just paraphrasing, and have we tested whether it works?
- Does our privacy policy account for inferred data, not just collected data?
- Sources
03 · SEO / Structured Data
3. Google adds sale dates and category to Merchant listing structured data
- What
- Google extended Merchant listing structured data to support two things it did not properly handle before: sale duration (start and end dates) and product category. You now use validFrom, validThrough and priceValidUntil to signal an active sale price, tied to the saleprice_effective_date attribute in Merchant Center, and Product.category can carry both merchant-defined and Google-defined category information. Both feed classic organic listings and AI-driven shopping results. This is free product visibility, not an ads feature, so the effort-to-reward ratio is unusually good. When your on-page schema and feed agree, you become eligible for strikethrough pricing that shows the original price crossed out beside the sale price.
- When
- Flagged on 15 July 2026 via PPC News Feed coverage aligning the schema docs with existing Merchant Center behaviour.
- How it shifts discovery
- Strikethrough pricing does the persuading before the click, and clean Product.category is the sleeper: AI shopping surfaces lean on taxonomy when assembling comparison results, so vague categories mean poor grouping or no surfacing at all. This is a one-sprint job. Audit your Product schema, add validFrom and validThrough to any live or planned sale in correct ISO 8601 with the right timezone, and reconcile against saleprice_effective_date so feed and schema match.
- Questions to ask
- Do our sale windows in on-page schema match saleprice_effective_date in Merchant Center exactly, timezone included?
- Is Product.category populated cleanly, or are we relying on Google to guess?
- Are we eligible for strikethrough treatment on our current live sales?
- Sources
04 · Paid / PPC
4. Performance Max gets a Theme Library for structured, testable asset groups
- What
- Google rolled out a Theme Library for Performance Max that lets teams generate themed asset groups with AI assistance instead of piling every image, headline and description into one undifferentiated bucket. A theme is a container grouping creative around a single angle: a use case, audience segment, season or product line. Google then maps each themed asset group to intent and audience signals, so it serves the right creative to the right person rather than blending everything into an average. Because PMax optimises at the asset group level, one flat group meant one signal soup; multiple themed groups give the algorithm cleaner inputs and you cleaner, theme-level reporting.
- When
- Rollout reported by PPC News Feed (published on the site 15 July 2026).
- How it shifts discovery
- The real story is control: this turns PMax from an asset dump into something you can structure and test like real segments. You can retire a weak theme without collapsing the campaign and scale a strong one with confidence. Audit your current asset groups this week, define themes by intent and audience rather than feed alone, and build one asset group per theme.
- Questions to ask
- Are we still running one flat asset group per PMax campaign and losing signal clarity?
- What intent and audience angles should define our themes, beyond the product feed?
- Can we now read theme-level performance to decide what to scale or cut?
- Sources
05 · Analytics / Ecommerce
5. The Shopify app migration that could wipe your product performance history
- What
- Reinstalling the Google & YouTube Shopify app before the migration deadline may rewrite your product IDs in Google Merchant Center during the move from the older Content API to the newer Merchant API. A product ID is the primary key linking a listing to its entire history: impressions, clicks, conversions and the machine learning signals bidding relies on. Rewrite the ID and Google treats the item as brand new, so months of history vanish, Merchant Center learning cold-starts, product-level attribution breaks and bidding turns volatile. Google Ads Liaison Ginny Marvin confirmed the team is investigating, but no fix has been announced. The trap is that the action triggering the reset (the reinstall or reconnect) is the exact action Google is nudging everyone to complete before the cut-off.
- When
- Migration deadline is 18 August 2026. The risk window is the reinstall or reconnect step ahead of that date.
- How it shifts discovery
- Plumbing changes rarely announce themselves as strategy changes, but they hit performance all the same. Before you touch anything, pull a full product performance export from Google Ads and Merchant Center with current IDs mapped to SKUs, and record your current ID format. Test on a staging or low-risk store first, and check the official migration guidance before reconnecting your revenue driver.
- Questions to ask
- Have we exported product performance and mapped current IDs to SKUs before any reconnect?
- Can we test the migration on a low-risk store before touching our main revenue store?
- Do we know what ID format Shopify sends now, so we can detect a change after?
- Sources
06 · AI Search / Ecommerce
6. Bing turns its Shopping results into the storefront with a product overlay
- What
- Bing is testing a product detail overlay that surfaces images, descriptions, retailer pricing and price insights inside the results page, with sponsored Shopping results at the top. When a shopper clicks a product, an overlay opens showing imagery, descriptions, prices from multiple retailers and a price comparison chart, all inside the SERP, so the shopper can research and compare without visiting your site. Bing pulls structured product data and renders it as a self-contained shopping experience, which collapses the funnel: the overlay is the product page now. The quality of your feed decides whether you win the comparison, not the design of a page the shopper may never reach.
- When
- Overlay surfaced in the wild on 11 July 2026. It is a test, not a global rollout, following a bigger Shopping carousel in March and underlined product titles in February.
- How it shifts discovery
- This mirrors the zero-click shift already reshaping Google and makes feed quality a core discovery skill. Audit your product feed as if it were your homepage, fix pricing accuracy first because a stale or higher price loses the sale before a human sees your brand, strengthen product-page structured data so both Bing and Google render you cleanly, and reframe measurement to value in-SERP impressions even when the click never lands.
- Questions to ask
- Is our feed pricing accurate enough to win a live comparison chart?
- Are our feed titles, descriptions and imagery merchandising-grade, or still backend plumbing?
- How do we measure visibility inside overlays when the click never reaches us?
- Sources
07 · AI Search / Audience
7. OpenAI Signals: ChatGPT's user base is now majority non-English and majority female
- What
- OpenAI published its Signals data showing users predominantly using a language other than English now represent over half of active users, and usage by people with typically-female names makes up most usage globally. Signals measures aggregated usage across Individual plans (Free, Go, Plus and Pro). Three shifts stand out: six months after signing up, users sent 50% more messages per day and doubled the distinct tasks they tried across 53 task categories; the fastest relative growth since July 2023 has been in Africa and Asia and in lower-HDI countries, helped by low-cost Free and Go plans; and Spanish, Portuguese and Arabic lead the non-English pack, while Uzbek, Kazakh and Burmese posted the largest share increases. Brazil, Colombia, Poland and Namibia rank among countries where messages from users with typically feminine names most exceed masculine ones.
- When
- Published 30 June 2026.
- How it shifts discovery
- The centre of gravity for AI-assisted discovery has moved, so a GEO strategy built for an English-speaking, male-skewed North American core is optimising for a shrinking slice. Engagement compounds and breadth widens, meaning each new task category is a place your brand can be surfaced or absent. Review which languages and markets your content and structured data actually serve, and prioritise non-English coverage in high-growth regions.
- Questions to ask
- Does our content serve the non-English, high-growth markets where ChatGPT usage is expanding fastest?
- Are we present across the range of task categories users mature into, not just search-style queries?
- Is our audience assumption still English-first and North American, and is that costing us?
- Sources
08 · Paid / Transparency
8. Google now labels AI-made ads in My Ad Center
- What
- Google has started showing users when an ad's creative was made or edited with generative AI, via a new 'How this ad was made' panel inside My Ad Center across Search, YouTube and Discover. It is the first consumer-facing transparency layer on paid creative and sits alongside the existing Like, Block and Report controls. Two mechanisms sit behind it: Google automatically labels ads built with its own AI tools, and advertisers using third-party AI can manually flag that usage, with Google saying using the labelling tools will be required. Non-AI creative gets no label.
- When
- Rollout confirmed 12 July 2026, appearing across Search, YouTube and Discover.
- How it shifts discovery
- The risk is not the label itself, it is inconsistency: if Google auto-labels a Google-made asset but your team forgets to flag an externally generated one, you look like you were hiding something. Audit your creative pipeline and record which assets touch AI and which tools, set a disclosure default that labels all AI-assisted external creative, tighten creative QA so fabricated features or off-brand imagery do not ship under a visible AI stamp, and brief legal and brand early. Before your next flight, write a one-page AI disclosure policy and add an 'AI used?' field to every creative brief.
- Questions to ask
- Which of our creative assets touch AI, and are we tracking the tool used?
- Do we have a consistent disclosure default, or will some assets get labelled and others missed?
- Where is our line between 'AI-assisted' and 'AI-generated', and has legal signed off?
- Sources
09 · AI Search / Operations
9. ChatGPT Work turns the chatbot into a worker you delegate to
- What
- OpenAI launched ChatGPT Work, an agent inside ChatGPT powered by Codex and GPT-5.6 that takes action across your apps and files, stays with a project for hours, and turns a goal into finished work. Instead of asking for an answer and copying it out, you describe an outcome and it produces the document, deck, analysis, site or report, drawing context from the apps and files you connect. The loop is goal, context, execution and output: you state the outcome, it reads connected apps and files to ground its work, it takes actions across those tools holding the task for hours if needed, then hands back a template-matched artefact you review. GPT-5.6 is built to reason through complex multi-step tasks and match your templates and style.
- When
- Live from 9 July 2026 on web and mobile for Pro, Enterprise and Edu, with Plus and Business following over the next few days. In the desktop app, Chat, Work and Codex are available on every plan including Free, globally on Windows and Mac.
- How it shifts discovery
- This is the shift from prompting a tool to delegating a workflow, and the obvious fits are repetitive builds that eat analyst hours: weekly reporting, campaign briefs, audience research syntheses and first-draft landing pages. Agents are only as good as the context you feed them, so lock governance down before deployment: connect read-only or sandbox environments before granting write access, require human sign-off on anything that publishes, spends or emails, and keep an audit trail of what the agent did and where its inputs came from.
- Questions to ask
- Which repetitive builds could we safely delegate, and which need human sign-off?
- Have we scoped agent access to read-only or sandbox before granting write access to ad accounts or CMS?
- Do we have an audit trail so agent output is traceable to its inputs?
- Sources
10 · Analytics
10. Google Analytics: fine-tune conversion lookback windows for accurate attribution
- What
- Google Analytics is reminding teams that the buyer journey does not stop at 30 days and that ad interactions can spark actions days or weeks later. Fine-tuning your conversion lookback windows in GA4 ensures attribution credit lands where it is actually earned, rather than being cut off by a default window that undercounts delayed conversions. The guidance points to the official GA4 support documentation on configuring lookback windows.
- When
- Posted by @googleanalytics on 15 July 2026.
- How it shifts discovery
- If your default lookback window is too short, you are systematically under-crediting channels that drive delayed conversions, which distorts budget decisions. Review your conversion lookback settings against your real sales cycle, extend the window where the journey genuinely runs longer, and re-check attributed channel performance once adjusted so your budget follows the truth rather than an arbitrary default.
- Questions to ask
- Does our lookback window match our actual sales cycle, or is it stuck on a default?
- Which channels are we under-crediting because conversions land outside the window?
- Have we re-checked channel performance after adjusting the window before reallocating budget?
- Sources
Key takeaways
What to walk away with this week
The EU's DMA decision forces Google to share anonymised Search data with rivals and AI chatbots, the first real crack in its query-data moat: baseline your non-Google discovery metrics now.
LLMs can infer personal attributes people never disclosed, and people block it only 28% of the time: treat inferred data as a real privacy risk and sanitise by abstraction, not paraphrasing.
Feed quality is now a core discovery skill: Bing's overlay and Google's new sale-date and category schema mean your feed, not your landing page, wins the comparison.
Guard your data plumbing: the Shopify app migration (deadline 18 August 2026) can wipe product IDs and performance history, so export and test before you reconnect.
AI is now visible and delegable: Google labels AI-made ads from 12 July, and ChatGPT Work acts across your apps, so set disclosure policy and agent governance before, not after.
