SEO Feed

What to do next in SEO, from people doing the work.

Practical takeaways from experienced SEOs, summarized and linked back to the original post.

llms.txt got 2 LLM visits in a 30-day test, so learn how LLMs work before chasing hacks

Rota argues that most GEO and AEO advice on LinkedIn is unreliable, and that people adopt these tactics without understanding how LLMs work. Rota points to an earlier test where LLMs requested llms.txt twice in 30 days compared with 424 requests for the HTML sitemap, a result that casts doubt on whether the file helps visibility. The advice is to build on fundamentals rather than short-term hacks.

What to do

  • Do not rely on llms.txt to improve AI visibility.
  • Learn how to influence the sources LLMs cite.
  • Learn how to measure brand visibility correctly.
  • Work out where AI agents can and cannot help.

Evidence: The author's own 30-day test, described only briefly in this post. It measures visits to the file, not its effect on visibility. The post is tagged as a Profound partner post and promotes that company's courses.

Google AI Mode answers about your brand can draw on Knowledge Graph data with no link back to you

Dan Petrovic at DEJAN examined the code behind Google AI Mode and found it returns full biography results for people and companies straight from the Knowledge Graph, with no URL pointing to their sites. The data is richer and more current than Google's public Knowledge Graph API returns, and the set of result types and providers is still growing. A "Relevant webpages" field ties pages to your entity and feeds these answers, so if it is wrong, AI Mode repeats the error with no link for readers to check.

What to do

  • Treat influencing your Knowledge Graph entry as a long-term program, not a quick fix.
  • Work across on-page, off-page, social, PR and structured data signals.
  • Open the three-dot menu on an AI Mode result to see what Google says about the result.

Evidence: Code-level findings published by Dan Petrovic at DEJAN and relayed by the author, who says the three-dot menu in AI Mode reflects them.

Your XML sitemap should list only the URLs you want crawled and indexed

After reviewing a post-migration sitemap full of 404s, 301 redirects and irrelevant URLs, Rota argues that a sitemap should not contain every URL on a site. It should hold only indexable, canonical pages that return a 200 status, with accurate lastmod values, and it should update automatically as URLs change. Rota also warns that a sitemap is no substitute for internal linking to your important pages.

What to do

  • Limit the sitemap to indexable, canonical URLs that return a 200 status code, with an accurate lastmod.
  • Remove redirects, 4xx URLs, noindex or robots-blocked URLs, tracking URLs, login pages, thank-you pages and private URLs.
  • Make the sitemap update automatically as URLs are added or removed.
  • On large sites, use a sitemap index; the limit is 50k URLs per sitemap, though the author aims for 20-30k.
  • After a migration, crawl the sitemap to find broken URLs, redirects and non-indexable pages.
  • Use a tool like IndexNow to notify supported search engines like Bing about published, updated or deleted content.

Evidence: The author's experience, prompted by one recent post-migration sitemap review. The post gives no test data.

Adobe's retail data shows AI-referred visitors now convert 90% better than other traffic

Adobe's retail data shows AI-referred traffic has reversed from converting 9% worse in August 2025 to 90% better now, with 83% higher revenue per visit and 71% year-over-year traffic growth in September 2026. Toth reads this as visitors arriving further along in their decision because they did their research inside the LLM. Adobe also found that more than a third of homepage content is invisible to AI systems, and that average US retail homepages scored 63% on its Citation Readability Score.

What to do

  • Track AI referrals as a separate segment in your analytics.
  • Rewrite product pages with plain-text specs, prices and buying context that AI can parse.
  • Strengthen FAQ, buying guide and return policy pages, the content types Adobe found AI favors most.

Evidence: Adobe Digital Economy Index (Oct 29, 2026), based on more than 1 trillion visits across 18 retail categories. It covers retail only, and the explanation for why these visitors convert better is the author's.

Most AI citations vanish within days, but many return and brand mentions last longer

Ahrefs tracked 37,500 AI citations across 5 platforms for 20 days. About 60% disappeared by day 2 in AI Mode and ChatGPT, yet 75% of lost citations returned within 20 days. Brand names stayed far longer than cited URLs, and pages ranking anywhere in Google were 7 times more likely to hold a citation for a week or more. "Best X" queries were the least stable, with 64-76% of cited pages changing daily on ChatGPT and AI Mode.

What to do

  • Measure citations over longer periods instead of judging them day to day.
  • Track brand mentions alongside citation URLs.
  • Keep investing in traditional SEO, since Google rankings were the strongest predictor of citation longevity.

Evidence: An Ahrefs study of 37,500 citations over 20 days, reported by the author, who says the same churn shows up in the author's own tracking.

ChatGPT favored the product that published exact specs, even when the rival was just as good

Toth's team ran 20 product comparison questions for a consumer electronics client and captured the pages ChatGPT read. The competitor won 15 of 20 because it published exact ranges, named modes and named sensors, while the client only called the same feature "precise", leaving no number to quote. In the matchup where the competitor published fewer specs, the client won. Toth's reasoning is that without a link graph, ChatGPT leans heavily on first-party sources, so product pages have to carry the evidence.

What to do

  • Publish exact product facts such as range, weight, dimensions, battery life, heat-up time and what's in the box.
  • Put specs in a real HTML table.
  • Add Product schema and FAQ schema so each fact can be extracted in one pass.
  • Re-run the same prompts every sprint and track who ChatGPT quotes, not just who it mentions.

Evidence: The author's agency test: 20 buyer questions for one client, with captured retrieval. It covers a single brand and product category, and the explanation of why ChatGPT behaves this way is the author's interpretation.

AI Overviews now appear on most brand searches, not just informational queries

Toth pulls together figures from three data providers to challenge the idea that AI Overviews are limited to informational searches. SE Ranking puts them on 82.5% of US brand searches, DemandSphere saw brand coverage climb from 26% to over 80% during September, and seoClarity reports them on 62% of all queries. Toth concludes that Google is now writing an AI answer about your brand on most brand searches, while Search Console still offers no official click data for them.

What to do

  • Structure your content so it can be cited inside AI Overviews.
  • Use entity optimization to shape what AI answers say about your brand.
  • Plan on not getting official AI Overview click metrics in Search Console any time soon.

Evidence: Third-party data from SE Ranking, DemandSphere and seoClarity, plus a Google statement to Barry Schwartz about click data. The conclusions drawn from it are the author's own.

Locale page variants can take up several slots in ChatGPT's retrieval pool without earning more citations

In one captured ChatGPT session, a payroll vendor held 12 of 28 retrieved pages but earned only 2 citations, and three of those pages were the same product page in different locales. Toth's explanation is that the pipeline strips tracking parameters when de-duplicating but does not normalize locale prefixes, and nothing in the capture suggests it reads hreflang. Toth stresses that this does not prove the variants caused the low citation rate, and that brands should not change correct international SEO because of it.

What to do

  • Keep your international SEO setup as it is rather than working around this.
  • Treat this as a signal to watch, not proof that locale pages reduce citations.

Evidence: A single captured 3-turn ChatGPT session on one plan tier in one geography, which the author says is not a controlled study. A rerun would retrieve a different pool of pages.