SEO Recap covering September 10, 2026: Crawl-to-refer scrutiny and a 2026 ranking factors survey

Built by Stephanie Chung·9 min read·11 stories

Daily summary of what matters in SEO, GEO, AEO, and AI search generated with Claude Code (beware of hallucinations)

Generative Engine Optimization (GEO) & Answer Engine Optimization (AEO)

Anthropic's crawl-to-refer ratio has been quoted 7 different ways
Source: Search Engine Journal

  • Anthropic's crawl-to-refer ratio has appeared as 70,900 to one, 38,000 to one, 23,951 to one, 11,122 to one, 10,300 to one, 4,580 to one, and 2,237 to one, all attributed to Cloudflare within ~13 months, with 2 claiming the same month and differing by ~17 times.
  • Crawl-to-refer compares how many pages a platform's crawlers fetch against how many visitors it sends back. Publishers use it to decide which AI crawlers to allow or block, and marketers cite it to argue AI referral traffic is not worth pursuing.
  • The spread comes from 4 undisclosed choices baked into the figure:
    • The measurement window (Cloudflare saw Google's ratio move 19.4% week over week from a crawl scheduling change)
    • Which bot got counted (training crawlers and user-request crawlers are aggregated under one platform name)
    • Whose sites were sampled (panel composition alone roughly doubled one platform's ratio)
    • Whether referrals announced themselves (Claude's native app sends no Referer header, so referrals are undercounted by an amount Cloudflare says it cannot size)
  • When handed any metric, ask what period it covers, what got grouped to produce it, and where collection stopped. A figure without its window, grouping, and boundary looks usable but is not.

How Freshpet built a GEO program after strong SEO fell short
Source: Search Engine Journal

  • Freshpet ranked well in traditional search but stayed nearly invisible when buyers asked AI assistants for fresh pet food recommendations, showing AI visibility is not an automatic extension of rankings.
  • Intero Digital's framework examines 4 areas: relevance, authority, structure, and engagement, and works best when onsite content, SEO, earned media, and social reinforce the same priority topics.
  • Freshpet rewrote magazine-style content into focused sections with descriptive headings, short paragraphs, lists, and schema so LLMs could extract facts, and found product reviews and Q&A delivered via JavaScript were invisible to crawlers.
  • Start with a set of at least 20-30 prompts run across ChatGPT, Gemini, and Claude, recording whether the brand appears and which sources are cited. Freshpet tracks 475 prompts grouped by topic and prioritizes fixes closest to revenue.

How local businesses can build visibility in AI search
Source: Search Engine Journal

  • AI visibility for a local business means being the company an assistant recommends when someone describes what they want, and models pull from listings, review sites, Reddit threads, city forums, and curated local lists well beyond the map pack.
  • Google changed its Maps content policy in April to bar merchants from requesting a set number of reviews or reviews mentioning specifics like a staff member's name, ending staff leaderboards and "mention me" cards.
  • Keep name, address, phone, and hours consistent across platforms and keep the site well-organized, since models now treat the site as a source more than a place customers read.
  • When a business ranks first on Google but is absent from AI answers, run prompts across engines to see which URLs are cited, then check whether AI crawlers can access the site, since outdated robots.txt rules sometimes block them.

AI in Search / AI Overviews

Google DeepMind researches an LLM to replace two-stage search ranking
Source: Search Engine Journal

  • A paper from Google DeepMind, UMass Amherst, and UT Austin proposes Autoregressive Ranking (ARR), a single LLM that outputs a ranked document list in place of the current Dual Encoder plus Cross Encoder architecture.
  • Dual Encoders retrieve candidates cheaply but rank imprecisely, while Cross Encoders rank better but are too expensive for large-scale retrieval. ARR aims to bridge both.
  • The team trained ARR with a method called SToICaL (Simple Token-Item Calibrated Loss), using item-level reweighting and prefix-tree marginalization to weight tokens toward higher-ranked documents.
  • In WordNet tests ARR matched the Cross Encoder and beat the Dual Encoder, but in a shopping-search test one variant got worse at ranking the single most relevant result first. Results are largely theoretical, not proven in live search.

Le Monde says AI Overviews haven't hurt its audience so far
Source: Search Engine Journal

  • Le Monde CEO Louis Dreyfus said the audience hasn't dropped abruptly since AI Overviews launched in France, contrasting with press alliance APIG's filing that cites Arcom's estimate of 33-38% search traffic loss in European markets.
  • Le Monde in English, read in AI Overviews markets, grew 7% in the first half of 2026 versus 2025, and apps generated 36% of group traffic in August, insulating it from web-only declines.
  • Dreyfus said news content stayed relatively stable while service sectors like health, food, and automotive may have seen declines up to 50%, based on conversations with publishers rather than published data.
  • ChatGPT visitors convert to subscribers at rates 26 times higher than Facebook and 260 times higher than Discover, and Le Monde holds content agreements with OpenAI, Perplexity, and Meta.

Technical SEO

Mueller says the old "position 1-10" is hard to map in Search Console
Source: Search Engine Roundtable

  • John Mueller said search results pages now have so many ways for users to interact that the old "position 1-10" is hard to map or make useful for site owners.
  • Generative AI impressions are tracked as a block, like other search features, and are not separated out in the Gen-AI performance report.
  • Mueller pointed to the help doc found by searching "impressions clicks search console" and invited feedback on what position tracking would actually be useful.

Organic Search & Algorithm Updates

2026 Google ranking factors survey from Cyrus and Dawn Shepard
Source: Search Engine Roundtable

  • Over 100 SEO experts submitted 13,665 ranking-factor data points for the Zyppy survey, the first such poll in a while.
  • Top tier factors (over 45%): content relevance (57.1%), backlinks (54.8%), and content quality (47.6%).
  • Mid tier (25-40%): trust/authority (36.5%), behavioral click signals (29.4%), and brand presence (27.0%).
  • Foundational and specialized factors under 20%: user satisfaction, technical SEO, topical authority, and internal linking.

Content marketing results hit a 12-year low despite 92% AI adoption
Source: Search Engine Land

  • Orbit Media's 2026 survey of 1,042 content marketers found just 14% said their blogs delivered strong results, 6 points below the prior 12-year low and nearly half the 26% reported in 2022.
  • AI adoption reached 92.4% but showed no relationship with stronger results. It cut average writing time to 3 hours 20 minutes per post from over 4 hours in 2022, saving ~50 hours a year.
  • The decline tracks marketers abandoning proven tactics: influencer collaboration, original research, keyword research, paid promotion, formal human editing, and consistent analytics use.
  • Influencer collaboration had the strongest tie to success, with regular expert collaborators 2.6 times more likely to report strong results, yet adoption fell from 25% in 2017 to 7% this year. Marketers who did keyword research also reported stronger outcomes.

When SEO best practices become dogma
Source: Search Engine Land

  • Many best practices (one H1, title tag character limits, keyword in title and H1, all-green Core Web Vitals, schema) began as reasonable observations but became checklist items once the reasoning got lost.
  • Something can be technically true without being worth prioritizing. A slow high-traffic conversion page merits work, while a low-traffic resource page missing a benchmark by a fraction of a second may not.
  • Issue counts don't equal impact. 10,000 missing alt attributes on legacy images can be trivial, while one canonicalization or rendering problem on a key template can affect thousands of URLs.
  • Ask better questions before acting:
    • What evidence suggests we should change what we're doing?
    • What problem does this actually solve?
    • How much of the site does the issue affect?
    • What happens if we don't fix it?
    • Does the reason it worked elsewhere apply here?

Losing a client to AI software and the cost of commoditizing SEO judgment
Source: Search Engine Land

  • A consultant lost a prospect to an AI tool promising automated keyword research, content briefs, technical audits, and execution for a fraction of a monthly retainer.
  • AI can surface keywords, generate briefs, run technical checks, and produce content at scale, but it can't judge whether that traffic attracts the right customer, whether to consolidate rather than publish another page, or which of 100 flagged issues actually matter.
  • Reducing SEO to tasks (keyword research, briefs, audits, reports, pages per month) makes a consultant easy to compare against software on price and volume, obscuring that clients pay for judgment and ownership of the recommendation.
  • Turning 5 thoughtful articles into 50 mediocre ones at half the cost is not automatically an efficiency win when the output can't earn attention, trust, links, or revenue.

Using Claude to run a stronger CRO audit
Source: Search Engine Land

  • Define the conversion the audit should improve before uploading data. A GA4 key event only marks importance in reporting, not that it fires correctly or represents a qualified outcome.
  • Keep a one-page audit brief in a Claude Project covering primary conversion, quality measure, measurement source, date ranges, scope, recent site changes, known limitations, and business constraints.
  • Add standing rules so Claude separates observed facts from hypotheses, avoids claiming causation from correlations or screenshots, cites the supporting file for each finding, assigns a confidence level, and flags when evidence is insufficient.
  • Provide data via curated exports or a read-only MCP connection, with least-privilege access scoped to one property and a named event. Give Claude bounded tasks rather than "run a CRO audit."
Published on
geoai searchtechnical seoorganic searchdigest