Writing
Deep dives, clustered by argument.
01 · Pillar
The Feed Is the Storefront
The structured product feed is no longer a Shopping input; it is the surface AI rebuilds commerce around. Control is completeness, accuracy, and representation.
Google's Universal Cart tells the retailer nothing until the shopper clicks transferGoogle's Universal Cart keeps the shopper's cart on Google. Its developer guide says retailers get one request, at transfer, and nothing while items are saved.The Feed Is the StorefrontWebMCP Checkout: Shopify Asks Agents to Sign Who They AreShopify's WebMCP checkout asks agents to sign requests with Web Bot Auth. Amazon said Meta's Muse never identified itself. Order visibility is unanswered.The Feed Is the StorefrontDo agentic checkout protocols require idempotency keys?Do agentic checkout protocols require idempotency? OpenAI's ACP has since April 2026. Google's UCP requires the key in its schema, and its prose says SHOULD.The Feed Is the StorefrontIs My Product Catalog Shared Automatically? Two CasesShopify shares products to Meta by default. Microsoft publishes your catalog when you accept an analytics tool's terms. Neither is a syndication decision anyone made.The Feed Is the StorefrontHow Many GTINs Can a Product Have? Google Accepts 10Shopify variants now store 20 barcodes. Google accepts up to 10 per product through every submission path. The export forwards the first one, and only the first.The Feed Is the StorefrontUCP Spec Breaking Changes: What Merchants Must FixThe Universal Commerce Protocol's August release rewrites fulfillment, consent and signing keys. Four releases in, breaking changes are the pattern, and merchants track two rival specs.The Feed Is the StorefrontAnthropic Commerce Agents: You Own the RankingAnthropic's commerce agents blueprint ships no payment code, no feed format, and no ranking. The method deciding what a shopper finds is one you write yourself.The Feed Is the StorefrontAggregator unit for Europe: a comparison shopping featureOn product queries, every route Google gives for its European aggregator unit points into the Comparison Shopping Services help center.The Feed Is the StorefrontAI Catalog Enrichment Ships Unverified AttributesTRACE's verifier has four verdicts and two of them publish. UNVERIFIED means it could not confirm the value for your exact product, and it ships like PASS.The Feed Is the StorefrontProduct Feed vs. Availability Record for AI AgentsA product feed records what you sell. A separate availability record, often held by a booking intermediary, decides what an agent can actually book.The Feed Is the StorefrontIncentivized Reviews Have No Schema.org Disclosure FieldGoogle bars undisclosed incentivized reviews from structured data. Schema.org's 9 Review properties have no field to carry that disclosure.The Feed Is the StorefrontDeal Ends badge runs on sale_price_effective_dateGoogle's new Deal Ends countdown badge is generated from sale_price_effective_date, a feed field with no schema.org equivalent and no automatic safety net.The Feed Is the StorefrontGoogle Tests AI-Written Shopping Ad DescriptionsGoogle is testing AI-generated Shopping ad descriptions. Bing and ChatGPT Ads are testing feed-rendered pricing and cards, and none of the three has launched.The Feed Is the StorefrontEligible-not-distributed gap in Google Merchant CenterMerchant Center approval confirms a feed passed checks. Whether Google actually shows the product depends on feed data quality and per-country attributes.The Feed Is the StorefrontContent API sunset Aug 18: check your storeGoogle sunsets the Content API for Shopping on August 18, 2026. Here is how to check in five minutes whether your store's ads and free listings are at risk.The Feed Is the StorefrontGoogle's merchant-listing schema update: a convergence taxGoogle added category and sale-duration properties to merchant listing structured data. For most merchants it adds reconciliation work, not a ranking advantage.The Feed Is the StorefrontShopify Catalog: build vs. syndicate for AI commerceShopify Catalog is positioned as the agentic product-data substrate, and a new Agentic plan lets non-Shopify merchants sync in and sell across AI channels.The Feed Is the StorefrontGoogle's patent: AI landing pages from your feedA Google patent describes scoring landing pages and substituting AI-generated alternatives built from the retailer's own product data.The Feed Is the StorefrontOpenAI's feed pivot: two pipelines for ChatGPTNative in-chat checkout retreated to apps while ACP expanded to organic, feed-driven discovery. Feed hygiene is the single lever for ChatGPT product visibility.The Feed Is the StorefrontConversational ads: the model writes copy from your feedAmazon, Google, and OpenAI are converging on one mechanic: AI generates ad copy at query time from the merchant's product feed, not advertiser-written creative.The Feed Is the Storefront
02 · Pillar
How Machines Read the Web
Agents and crawlers read structure, not pixels. Semantic HTML, server-rendered schema, and embeddable item data decide what gets understood and surfaced.
For about a week, ChatGPT stopped tagging its citation links, so GA4 filed those clicks as DirectFrom GPT-6 on October 7, 2026 until a fix reported October 15, many ChatGPT citation links lost their utm_source tag, so GA4 filed the clicks as Direct.How Machines Read the WebSearch Console's generative AI control: the closest property that changed it decides the settingIn Search Console's generative AI control, the closest parent property that stopped inheriting decides the setting, so a child property can override the domain.How Machines Read the WebDoes Google Pay Publishers for AI Overviews? Only in a PilotDoes Google pay publishers for AI Overviews? Only in a Search Console pilot, confirmed to one outlet, that pays for an event Google's AI report never counts.How Machines Read the WebLighthouse's AI catalog audit rewards publishing nothingLighthouse 13.5's AI catalog audit scored this site 4 of 4 with no catalog, and 4 of 5 with the catalog a content site would write first.How Machines Read the WebSearch Console AI Report: Impressions, No ClicksGoogle's generative AI report shipped worldwide with impressions and no clicks. Three separate changes hit AI citations in the same two weeks, and the report cannot price any of them.How Machines Read the WebWhat agent-readiness scanners actually checkA public scanner rated this site level 4 of 4 for AI agents while three of its machine-facing claims were false. Structural checks cannot see a wrong value.How Machines Read the WebGoogle's crawlers download JSON, they don't parse itGoogle now applies one pass of HTML unescaping to JSON-LD, and its crawlers do not parse JSON at all. A successful fetch says nothing about whether your data parsed.How Machines Read the WebGoogle goto Redirects and URL-Level AttributionGoogle confirmed google.com/goto redirects are rolling out. Domain-level rank tracking survives the change, and URL-level attribution is what degrades.How Machines Read the WebChatGPT answered 56% of queries without searchingAcross 270 alternatives queries, ChatGPT searched the live web on 44%. On the rest nothing you publish this quarter can change the answer.How Machines Read the WebBlocking AI crawlers did not stop AI citationsAround 75% of sites blocking OpenAI or Google AI bots were still cited. Their own table shows why: nobody blocked the retrieval bot, only the training bots.How Machines Read the WebCitation half-life and your content refresh queueA vendor now plots per-URL citation decay and says half of cited content is under 13 weeks old. The metric is useful. The figure has no published method.How Machines Read the WebCloudflare's AI agent checklist, tested on 15 retailersCloudflare published a real, checkable AI-agent specification. A 15-retailer test shows why the score placed beside it should not be trusted.How Machines Read the WebWhy most AI-driven site visits skip the referral tagA panel study finds AI mentions lift site visits 1.5 to 2.5x, but for ChatGPT, only about 2.5% of those visits carry a trackable referral tag.How Machines Read the WebGoogle Tests Sign-In Wall on Search Results After SerpApiGoogle is testing a sign-in wall on deep search results weeks after a DMCA loss to SerpApi, swapping an anonymous CAPTCHA for an identity check.How Machines Read the WebWhen an AI agent logs in, analytics see a humanThree AI agents now log into accounts and complete purchases, and this site's own crawler log proves no user-agent bucket can catch them.How Machines Read the WebAI Text Watermarking: What It Can and Can't ProveAnthropic now watermarks Claude's text. Google has watermarked Gemini's since May 2024 with no ranking signal to show for it. Here is what the mark proves.How Machines Read the WebCrawl Budget 304s: Why Most Retail Sites Still FailGoogle's crawl-budget guide now recommends serving 304 responses, but testing 13 live retail homepages found 11 still send the full page every time.How Machines Read the WebRobots.txt Disallow vs Noindex: What Each One Actually BlocksGoogle indexed Claude's blocked share pages via inbound links, then said noindex needs crawl access too. Same mechanism, opposite failure.How Machines Read the WebValidate Fix vs. Request Indexing in Search ConsoleValidate Fix samples a few pages before recrawling the rest, which only pays off for batch issues affecting many URLs.How Machines Read the WebChatGPT source selection: the editable policy behind itChatGPT's source selection runs on an editable server-side policy that OpenAI can rewrite at will, and the underlying fields already changed twice in July 2026.How Machines Read the WebGoogle's billions of clicks claim vs. the traffic dataGoogle says AI Search sends billions of clicks weekly and calls the number stable, but Cloudflare and Wikipedia both report reproducible traffic declines.How Machines Read the WebAnnotated Page Content (APC): what Gemini reads in ChromeChrome builds a structured, typed node tree from a page's render tree and hands that to Gemini when a tab is shared, per Dejan AI's teardown.How Machines Read the WebDoes AI prefer Reddit? OpenAI rejects it, Google keeps itOpenAI rejects 99.4% of the Reddit pages it retrieves, Anthropic cites zero, and only Google keeps Reddit near its baseline citation rate.How Machines Read the WebShopify's Hreflang Toggle Doesn't Fix the Real RiskShopify's new toggle turns off auto-generated hreflang tags, but it can't verify your replacement is correct or see CDN-level conflicts.How Machines Read the WebHow Brave Search finds pages (and cites via Claude)Brave has no Search Console, no sitemap ingestion, and no IndexNow. It rides on Googlebot-crawlability, and Claude's web search runs on Brave's index, not Bing's.How Machines Read the WebBot-Challenge Screens Can Get Your Real Page DeindexedGoogle's John Mueller confirmed bot-challenge screens served to Googlebot with a 200 status get indexed, and dedup can hand your canonical to a stranger's copy.How Machines Read the WebGoogle's nearest-seed patent: what it says vs what leakedA 2006 Google patent ranks pages by distance from a few trusted seed sites. It expires in 2026, but the leak shows the idea survived as PageRank-NearestSeeds.How Machines Read the WebEU AI Act Article 50: AI content disclosure rulesThe EU AI Act's Article 50 makes AI-content disclosure enforceable on August 2, 2026. Here is what gets marked, who has to disclose, and what it means for search.How Machines Read the WebWhat Is RAG (and When Not to Use It)RAG lets an AI look answers up instead of guessing from training. Here is how it works, what you can build with it, and when a simpler tool wins.How Machines Read the WebDoes Google use S-BERT to detect AI content? What the paper saysGoogle's S-CTS paper targets coordinated AI-slop channels on video platforms, not Search rankings. The real lesson is coordinated footprints, not your pages.How Machines Read the WebAI usage is a demand signal, not a jobs reportAnthropic's June 2026 Economic Index shows the unit of demand shifting from keyword to artifact. Read as a demand signal, it maps which intents AI is absorbing.How Machines Read the WebClaude Sonnet 5 vs Opus 4.8, and the agent cost curveClaude Sonnet 5 matches or comes within two points of Opus 4.8 on most agentic benchmarks, at a fraction of the price. For commerce, the story is the cost curve.How Machines Read the WebAgents read the accessibility tree firstProduction browser agents like Playwright MCP read the accessibility tree before pixels. For e-commerce, clean semantic HTML and schema decide machine readability.How Machines Read the WebSoft tokens: single embeddings for item representationGoogle is testing 'soft tokens', single embeddings that replace multi-token item descriptions, for cheaper representation with cold-start implications.How Machines Read the WebGoogle's crawler split: SSR vs. CSR and Merchant CenterThe Shopping crawler does not render JavaScript; it validates price and availability. SSR vs. CSR is the real determinant of Merchant Center compliance.How Machines Read the Web
03 · Pillar
The Model Is the Dependency
Every model you build on is a dependency you do not control. It carries its own behavior, spends your context, and runs on the credentials you hand it.
A code comment is a claim, and a review agent with no comment check passed a dead rule in half my runsWith no comment check, a review agent passed a dead ignore rule in 3 of 6 isolated runs. Told to check comments, it caught both defects in 5 of 5.The Model Is the DependencyMultiple Claude Code sessions in one repository: what breaksClaude Code sessions sharing one repository can silently revoke each other's write permission. Three logged incidents and three checks that should prevent it.The Model Is the DependencyReduce Claude Code token usage: shorten the sessionTo reduce Claude Code token usage, shorten sessions. Re-reading the conversation was about 76% of my consumption, and instruction size was ruled out.The Model Is the DependencyHow Much Human Review Does an AI Agent Need?Snowflake routes agent-proposed data-governance changes through human approval. Perplexity checks in less. Both are consistent, and the rule that reconciles them is about reversibility.The Model Is the DependencyRecall, not knowledge, is the AI factuality bottleneckFrontier models encoded 95 to 98% of tested facts and failed to recall 26 to 34% of them. Missing and unreachable look identical from the outside.The Model Is the DependencySkills, Gems, and Loops: One AI Pattern, Five NamesClaude Skills, Google Gems, and AI loops all pull a procedure out of the live prompt and into a file the model loads on demand.The Model Is the DependencySub-Agent Delegation: The Real Cost Is ContextToken cost is the shallow reason to delegate to AI sub-agents; the real constraint is the orchestrator's context window, and the failures stay undocumented.The Model Is the DependencyCan You Trust an MCP Server? A Practical AuditAn MCP server runs with your credentials and reads tool text a model can act on. Here is the real trust surface, and a 10-minute audit to run today.The Model Is the DependencyWhy AI Keeps Writing 'Not X, But Y'RLHF mode collapse, markdown training data, and reward bias each explain part of AI's 'not X, but Y' tic. None explains all of it, and the evidence disagrees.The Model Is the DependencyOpenAI killed Atlas: the consolidation patternOpenAI retired the Atlas browser under nine months after launch. The pattern is deliberate consolidation into ChatGPT, not a company that lost its direction.The Model Is the DependencyThe J-space, explained: reading a model's unspoken thoughtsAnthropic's global-workspace paper found the J-space, the internal patterns Claude reasons with. The real win is reading what a model thinks but does not say.The Model Is the DependencyWhen agents get desperateAn interpretability finding for checkout reliability: a model stays composed on the surface while an internal 'desperate' pattern drives corner-cutting.The Model Is the Dependency
04 · Pillar
Crawl, Index, Rank
Three stages, in that order: a crawler fetches the page, an index stores it, a ranking system picks it. Most ranking problems are really crawl or index problems, and both are measurable on your own site.
Orphan pages on this site still get crawled: six articles with zero internal links drew the same bot traffic as linked onesSeven articles on this site cannot be reached by clicking from the homepage, yet the six with no inbound links drew bot requests at the same rate.Crawl, Index, RankSitemap Couldn't Fetch in Search Console: Real CausesSearch Console's 'Couldn't fetch' status can sit on a sitemap that loads fine. Google ties one cause to site quality, and no fix on the file touches it.Crawl, Index, RankVideoObject Schema Required Properties: Google's 3 FieldsGoogle requires three VideoObject properties and recommends 11. On an article with a YouTube embed, only three of the 11 still need a decision.Crawl, Index, RankWhat does mobile-first indexing mean now? Two end datesWhat does mobile-first indexing mean now? Google finished it October 31, 2023, ended its desktop fallback July 5, 2024, and still sends desktop Googlebot.Crawl, Index, RankMeta name googlebot-image: why Google ignores itGoogle accepts only googlebot and googlebot-news in a robots meta tag, so meta name googlebot-image does nothing, while robots.txt accepts it.Crawl, Index, RankDoes a Video Transcript Help SEO? An Audit of 60 VideosA crawler that reads your page cannot watch the video on it. I checked all 60 of mine and none ships the words spoken in them, which is the only form a machine can read.Crawl, Index, RankDoes llms.txt Do Anything? A Study and One Broken FileCommon Crawl found most llms.txt files are unedited template output. Mine passed every check and still told crawlers a section of my site did not exist.Crawl, Index, RankHow to Verify Googlebot Is Real: The Two-Step CheckAny program can type Googlebot into its user agent. Google's real test is a reverse DNS lookup confirmed by a forward one, and I ran it live against five addresses.Crawl, Index, RankHow to Detect a Crawl Trap From Server LogsMost of what looks like crawler waste is probing for files you never published. In my own logs it was 80.1%, and a keyword denylist put the same traffic at 21%.Crawl, Index, RankRedirect Headers: What Your Browser Hides From YouA browser resolves a relative Location header and caches whatever it is told. A crawler reads the raw headers, and this site's own 301 is missing one of them.Crawl, Index, Rank404 vs 410: Google Treats Them the Same. 429 Is Different.Google's own docs say all 4xx errors except 429 are treated the same, so a 410 does not remove a page faster than a 404. The code that changes crawling is 429.Crawl, Index, RankLog Analysis With No Server Logs: The Free-Plan MethodA static site has no access log to analyze. You can collect one yourself, and mine showed 72% of this site's crawler log was a single credential scanner.Crawl, Index, RankRobots.txt: The Longest Rule Wins, Not the FirstWhen an Allow and a Disallow both match a URL, Google picks the longer rule path. Seven of 19 retail sites ship exactly that conflict today.Crawl, Index, RankEvery Google Page Indexing Status, TranslatedGoogle's Page Indexing report has 18 statuses in three tiers. On this site, 105 of 245 discoverable URLs are engineered never to be indexed, and that is correct.Crawl, Index, Rank