steerlab

steerlab search and AI-answer rules (SEO, AEO, GEO)

How the research pages stay easy to find in search, to share as link previews, and to quote in AI answers. Read it before adding or changing a research note or the research page. Visual rules live in DESIGN.md; this file covers only findability.

• SEO (search engine optimisation): Google and Bing can find, read and rank the page.
• AEO (answer engine optimisation): ChatGPT, Claude, Perplexity and Gemini quote the page when asked a question it answers.
• GEO (generative engine optimisation): AI summaries such as Google’s AI Overviews use the page as a source.

Metadata comes from one file

• experiments/notes.json describes every note: name, title, description, summary (one sentence with the main numbers and the random control), published and modified (YYYY-MM-DD), image_alt, keywords, citations.
• python experiments/shared/build_meta.py writes, from it:
  • in each page’s <head>, between <!-- meta:start --> and <!-- meta:end -->: the canonical URL, Open Graph and Twitter tags, and JSON-LD (ScholarlyArticle per note; CollectionPage, ItemList and SoftwareSourceCode on the research page);
  • the <time> dates in each note’s eyebrow (“Updated …” when modified differs from published);
  • experiments/sitemap.xml and experiments/llms.txt;
  • each note’s 1200×630 link-preview image, report/figures/og.jpg, cropped from its cover.
• Never edit the meta:start block, the sitemap or llms.txt by hand. Change notes.json and rerun.
• When a note’s content changes, set its modified date, and the research page’s modified when its list changes.
• The links workflow runs build_meta.py --check, and scripts/check_links.py fails if a note has no notes.json entry.

Every note has

• One <h1> (the title), then a clean h2 → h3 order; <html lang="en">.
• A <title> and meta description that state the finding, not just the topic.
• An author line in the meta row (name, with links to GitHub and LinkedIn; the same links as sameAs in the JSON-LD, from author.same_as in notes.json) and the publish date in the eyebrow.
• “In short” under the meta row: 130 to 170 words that answer the note’s question on their own, with the main numbers and the random control. Start with the question (“Can activation steering override a system prompt?”) and answer it in the next sentence. This is the paragraph an assistant quotes whole. It may sit inside a collapsed <details class="key fold"> “Summary” (with Key results) so readers reach the story first: the text stays in the HTML, search and assistants read it, and notes.json still drives the metadata.
• Definitions in “X is…” form the first time a key term appears (activation steering, a direction, an SAE feature, a random-direction control).
• A “Questions and answers” section before the appendix: 4 to 6 questions people would type, as <h3> headings, each answered in its first sentence with a number where there is one. Plain HTML only.
• “How to cite” in the appendix: a BibTeX @misc block with the note’s URL.
• Text in the HTML, not built by JavaScript: crawlers do not run scripts. Interactive parts (the reply explorer, 3D scenes) add to the text; they never replace it.
• Alt text on every image that says what it shows, with the plotted values for charts.
• Descriptive link text: no “here” or “click here”.
• A cover JPEG under about 400 KB, with width and height set on every <img>.

Never

• FAQ or HowTo schema (HowTo is deprecated; FAQ rich results are for government and health sites).
• Dataset schema (retired from rich results in late 2025).
• Thin or near-duplicate pages written for keywords: one page per real experiment.
• Invented numbers: no made-up search volumes or citation results. Record only what was checked.
• FID in performance notes: Core Web Vitals use INP.

Site limits

• GitHub Pages serves a project site under /steerlab/, so robots.txt cannot be set (it only works at the domain root, which returns 404 and so allows every crawler, AI crawlers included).
• The sitemap and llms.txt live in experiments/, the folder they describe: a sitemap may only list URLs under its own path.

Off the page

• Verify the site in Google Search Console and Bing Webmaster Tools with an HTML meta tag, and submit https://albertlieyingadrian.github.io/steerlab/experiments/sitemap.xml.
• Share each new note on LinkedIn and X; AI summaries weigh mentions across platforms more than backlinks.
• Once the notes are indexed, check monthly whether ChatGPT, Claude, Perplexity and Gemini cite them for questions the notes answer, and record the result.

This site is open source. Improve this page.