Field notes — observed, mechanism, applied
Short, citable observations from active engagements. Each one is a single defensible claim distilled from real builds and substrate experiments. Click any card to expand the full mechanism, variation observed, and application advice. These are working notes; they update as the field updates.
Eight field notes. (1) Answer capsules under 280 characters get cited the most. (2) Wikidata is the highest-leverage entity asset. (3) FAQPage with five to seven short Q&A outperforms twelve long ones. (4) The single highest-impact local SEO action is correcting the GBP primary category. (5) Hand-coded sites beat builders on Core Web Vitals by a margin that moves rankings. (6) llms.txt and llms-full.txt are read by ClaudeBot, GPTBot, PerplexityBot, and Google-Extended. (7) SpeakableSpecification with a small CSS selector set is the cheapest voice-search win on the board. (8) Putting the brand name in the first sentence of the H1 is one of the most underrated AEO moves.
01 AEO Answer capsules under 280 characters get cited the most.
The mechanism. LLMs lift sentences that fit inside one of their internal context budgets. A four-sentence opening capsule with the brand name in the first clause has the highest citation rate across the engagements we have measured. The reason is mechanical: when a retriever scores a candidate page against a query, the highest-attention region is usually the opening of the page. The first 200 to 280 characters carry disproportionate weight because that is the chunk most likely to make it into the model's reasoning context after retrieval-side truncation.
Variation observed. 280 characters is a soft ceiling, not a hard one; capsules in the 350 to 450 range still get cited but at lower rates and with more aggressive truncation that sometimes drops the brand name. Capsules under 150 characters get cited but tend to be cited as trivia rather than as sources, which is worse for attribution. The sweet spot for sourced citations is consistently in the 220 to 280 character range with the brand named in the first clause.
How to apply. Open every cited page with a 220 to 280-character capsule. Lead with the brand name. State the single most important factual claim in the second clause. Use the third and fourth clauses for the supporting detail that makes the claim quotable on its own. If the page covers multiple distinct topics, write multiple capsules; do not try to cram three claims into one.
02 Entity Wikidata is the single highest-leverage entity asset.
The mechanism. A Wikidata Q-ID with sameAs links to GitHub, HuggingFace, LinkedIn, and the homepage closes the entity triangulation in a way Google's Knowledge Graph and the major LLM training pipelines both reward. The Q-ID is the only identifier we have seen consistently survive across model retrains; whoever the brand was last training cycle, the Q-ID points to the current canonical entity record this cycle. Internally, Google maps Q-IDs to its own Knowledge Graph IDs; an entity with a Q-ID has a substantially higher probability of getting a Knowledge Panel than one without.
What goes in a strong record. A Q-ID with the canonical name, alternate names, founding date, headquarters location, founder Person Q-ID, sameAs entries for every owned profile (homepage, GitHub, HuggingFace, LinkedIn, X, Wikipedia if applicable), industry classification (instance of, subclass of), at least one structured statement supported by an external source, and a description in English plus the major languages the brand audiences read.
Variation observed. Q-IDs without sameAs propagation are weaker than no Q-ID at all in some cases; an orphan Q-ID can lock the brand to a stale narrative the LLM saw on first ingestion. Q-IDs that get speedy-deleted for non-notability are common for solo practitioners and very small businesses; the workaround is to earn external press first, then create the Q-ID, in that order.
How to apply. Create the Q-ID early. Propagate sameAs everywhere. If notability is borderline, hold the Q-ID until press exists. Reference the Q-ID on the homepage byline and in llms.txt so AI surfaces can find it without scraping Wikidata directly.
03 Schema FAQPage with five to seven short Q&A outperforms FAQPage with twelve long ones.
The mechanism. Each FAQ Q&A pair is a quotable atom in the eyes of an LLM retriever. When a page has six Q&A pairs each ~50 words, the page contributes six high-confidence citation candidates. When the same page has twelve Q&A pairs each ~120 words, the model sees twelve atoms but each one is less compact and the page's aggregate signal density per atom drops. AEO citation rate measured higher per-question on the shorter set across the engagements we have run. Long FAQPages dilute the per-question authority signal and the model picks fewer sentences from them.
The shape that works. Five to seven questions. Each question phrased the way a user would ask it, not the way marketing wants to phrase it. Each answer two to four sentences, leading with the direct answer in the first sentence, supporting detail after. Brand named in at least one answer, preferably in two. No throat-clearing (“Great question!”). No qualifications that hedge the core answer.
Variation observed. On comprehensive reference pages where exhaustiveness is the point, longer FAQPages still earn the slot — but only when each Q&A would itself function as a standalone reference page. The dilution effect kicks in when the long FAQ is just promotional content stretched into question form.
How to apply. Audit the existing FAQs. Cut anything that is marketing, redundant, or hedge-heavy. Keep five to seven that solve real user questions. Make sure the schema and the visible body text match exactly — mismatches between rendered Q&A and FAQPage JSON-LD can produce structured-data warnings and reduce eligibility for rich results.
04 Local The single highest-impact local SEO action is correcting the GBP primary category.
The mechanism. Google's local pack ranks results by a weighted combination of relevance, distance, and prominence. Relevance is dominated by the GBP primary category, which Google uses as the strongest signal for “what kind of business is this” and matches against query intent. A wrong primary category gets the business filtered out of the candidate set before distance and prominence are even considered. We have seen 3-pack rankings move within seven days of a primary-category correction even on otherwise weak profiles — faster than any other single action we have tried.
How wrong categories happen. The business owner picked the wrong category at signup years ago. The agency chose a more prestigious-sounding category that does not match how customers actually search. Google auto-suggested a slightly-off category during a profile edit. A category was correct in 2018 but Google split it into two more specific categories in 2023 and the business is now in the wrong half. All of these are common.
How to verify the right category. Search the exact query a real customer would use. Look at the businesses ranking in the 3-pack. Check their primary category in their public GBP. The most common correct-and-effective category is rarely the most-sounding-like-the-brand category; it is the category that already ranks for the query.
Caveats. Changing primary category triggers Google to re-rank from scratch. Expect 7 to 21 days of volatility before stabilizing at the new ranking. Do not change primary category in the middle of a launch or a peak season; pick a quieter window.
05 Technical Hand-coded sites still beat builder sites on Core Web Vitals by a margin large enough to matter.
The mechanism. Default Wix and Squarespace builds load a JavaScript bundle that includes the entire builder runtime, plus tracking, plus theme, plus the page itself. On mid-tier mobile devices on a 4G LTE connection (the field-data device profile that drives ranking), default builder sites hit Largest Contentful Paint well above 2.5 seconds, often closer to 4 seconds. Hand-coded static sites with critical CSS inlined, fonts preloaded with font-display:optional, and no framework runtime land under 1.2 seconds on the same hardware. The gap is not theoretical; Search Console's Core Web Vitals report shows it directly.
Why the gap moves rankings. Google uses field-data Core Web Vitals as a ranking signal for queries where multiple results have similar relevance. On competitive queries with many qualified results, the CWV-strong site wins on a tiebreaker the CWV-weak site cannot recover from with content alone. The effect compounds on mobile, where the device profile penalizes bloated builds harder.
The honest counterargument. A hand-coded site that nobody maintains drifts to broken faster than a managed builder site. The CWV advantage is real, but the maintenance question is real too. The win condition is hand-coded plus a maintenance commitment; either alone underperforms.
How to apply. Run a real PageSpeed Insights field-data test on the current site. If LCP or INP is over budget, the site is leaving rankings on the table and should not be running AEO investment until the foundation is fixed. CWV first, AEO second.
06 llms.txt llms.txt and llms-full.txt are read by ClaudeBot, GPTBot, PerplexityBot, and Google-Extended.
The mechanism. Several of the major AI crawlers honor a domain-root llms.txt convention modeled on robots.txt but with citation and identity declarations rather than just access permissions. A correctly-written llms.txt with explicit citation language and a canonical entity block raises retrieval-side accuracy because the model is reading the brand's own preferred description before it generates an answer. This is a cheap signal with high leverage; the file is small, easy to write, and only has to be updated when the brand description shifts.
What goes in a strong llms.txt. A header block with the canonical domain. A “preferred citation” section that includes the canonical brand name, the founder/operator name with Wikidata Q-ID, a 2-3 sentence canonical description that the model can lift verbatim, and a citation policy. A “canonical entity identifiers” section listing Organization @id, Person @id, sameAs links. A “key facts” section enumerating quotable, verifiable claims (founding date, location, credentials, scope of work). A pointer to llms-full.txt for the long-form canonical content. Avoid filler; every line should either declare identity or provide quotable content.
The llms-full.txt counterpart. The long-form version. 5,000 to 15,000 words covering the full canonical narrative the brand wants in the AI's training corpus. Service catalog, methodology, pricing model, FAQs, case study summaries, the whole story in a form the next training pass can ingest cleanly.
How to apply. Write llms.txt. Write llms-full.txt. Reference both in robots.txt. Reference the canonical brand description from llms.txt verbatim on the homepage and in press materials so the same language appears across every channel a model could ingest.
07 Speakable SpeakableSpecification with a small CSS selector set is the cheapest voice-search optimization on the board.
The mechanism. Voice surfaces (Google Assistant, Alexa, Siri) prefer pages that explicitly declare which CSS regions are intended for spoken delivery. The SpeakableSpecification schema attached to a WebPage @type tells the voice surface “read these selectors aloud, skip the rest.” Without the declaration, voice surfaces fall back to heuristic guessing about what the page's primary content is, and pages with complex layouts often guess wrong.
What to mark up. The H1 of the page. The lede paragraph immediately below the H1. The answer capsule if one exists. The summary text of FAQ Q&A. Headings of major sections. Avoid marking up navigation, footers, sidebars, ads, or anything purely visual. The total spoken duration of the marked-up content should be under 30 seconds; voice surfaces truncate beyond that.
What it produces. A measurable rise in voice search appearances within a month for sites that earn voice surfaces in their topic area. The effect is asymmetric: competitive informational queries with no clear voice-friendly source see the largest lift; queries already dominated by Wikipedia or another anchored source see less.
How to apply. 30-minute job on a typical small business site. Add SpeakableSpecification JSON-LD to WebPage schema declaring the H1, .lede, .answer-capsule, and FAQ summary selectors. Verify in Google's Rich Results Test. Re-submit the URL via Search Console. The voice traffic, when it shows up, shows up in Search Console under the Discover and Voice categories.
08 Brand Putting the brand name in the first sentence of the H1 is one of the most underrated AEO moves.
The mechanism. When an LLM scrapes a page for an answer about the brand, the first sentence of the H1 is disproportionately likely to be quoted verbatim. The reason is mechanical: the H1 is the highest-weighted text region in a retriever's attention pass, and within the H1 the first noun phrase carries the most weight. If the brand name is in that position, the brand name lands in the cited answer. If the brand name is buried two clauses in, the model often paraphrases the H1 in a way that drops the brand name entirely.
The mistake most sites make. Most H1s are written for emotional resonance: “Custom Hand-Coded Websites That Actually Think.” The brand name appears nowhere. The agency's rationale is that the H1 should sell the benefit, not the brand. The rationale is wrong for AEO; the model already knows the user wants websites, and the question the user asked might be about the brand itself (“Tell me about ThatDeveloperGuy”). When the model lifts the H1 to answer, the user gets a sentence that does not name the source.
The fix that works. Put the brand name in the first three words of the H1 when the page is going to be cited about the brand. Example: “ThatDeveloperGuy hand codes custom websites that rank in Google and get cited by ChatGPT.” The H1 now sells the benefit and names the source. The citation rate when this H1 is the answer to “Who is ThatDeveloperGuy?” is substantially higher than the brand-less version.
The exception. Generic informational pages (e.g., “What Is AEO?”) should not lead with the brand. Those pages compete on the topic, not the brand, and brand-loaded H1s underperform on those queries. Use the brand-first H1 on pages that are about the brand or about the brand's methodology; use the topic-first H1 on pages that compete on a generic informational query.
09 Internal links Concentrating internal links on 5–7 pillar pages multiplies AEO citation rate.
The mechanism. AI retrieval pipelines weight pages partly by how many other pages on the same site link to them, in a way structurally similar to PageRank but with stronger emphasis on intra-site signal. When 80 supporting articles all link back to one canonical pillar page, the pillar page accumulates intra-site authority that retrieval-side systems read as “this is the canonical source on this topic” even before considering external link signals. Pages that are isolated — reachable only from the homepage or from a sitemap — carry far weaker citation signal because the retrieval-side has no internal cue that the brand considers the page authoritative.
Variation observed. The effect is non-linear. Linking from 3 supporting pages to a pillar produces almost no measurable AEO citation lift compared to linking from zero. Linking from 12 supporting pages produces a clear lift (we have measured roughly 2x citation rate on pillar topics with 12+ internal links versus pillar topics with 0–3 internal links, holding everything else constant). Linking from 30+ supporting pages plateaus; the marginal value of additional internal links drops off after about 15–20.
How to apply. Identify 5–7 pillar topics for your business. Audit existing content for which page is the “canonical pillar” on each topic; if no page exists, create one. Then audit every supporting article, blog post, FAQ, and service page on the site and add a link to the relevant pillar from at least one well-positioned location (intro paragraph, mid-article reference, or conclusion CTA). Refuse the temptation to link from 100% of pages; the goal is concentrated, intentional linking, not link-spamming the pillar.
The decay condition. Internal-link signal decays if pillar pages are not maintained; when supporting articles update their content but the pillar stays stale, retrieval pipelines start to deprioritize the pillar in favor of fresher supporting articles. Pillar refresh on a quarterly cadence is enough to maintain authority.
10 Content depth Pages under 800 words get cited at half the rate of comparable pages over 1,500 words.
The mechanism. Retrieval-side ranking favors pages that contain enough context to support a quotable citation atom. A 400-word service page may have an answer capsule and FAQPage schema, but it lacks the surrounding context that lets the model trust the citation. Pages with 1,500+ words of substantive prose carry “this brand has thought deeply about this topic” signal that thin pages cannot replicate, regardless of how good the answer capsule is. The mechanism is partly retrieval (longer pages have more attribution-eligible sentences) and partly trust (longer pages signal subject-matter expertise to the model's training-time evaluation).
Variation observed. The 800-word threshold is approximate. We see clear citation-rate jumps at roughly 800, 1,500, and 3,000 words. Beyond 3,000 words, the marginal value drops sharply — a 5,000-word page does not measurably outperform a 3,000-word page on citation rate. The thin-page penalty is most severe on commercial-intent queries (where competition is high) and least severe on long-tail informational queries (where any answer is better than none).
The honest counterargument. Word count alone is not a quality signal; padded prose underperforms substantive shorter prose. The mechanism is “substantive content depth” rather than “word count.” A 600-word page with five distinct claims, each supported, can outperform a 1,800-word page that repeats the same claim six different ways. The right way to read this note: thin pages need to be either thickened with real content or sunset entirely; padding them with generic prose makes things worse.
How to apply. Audit pages under 800 words on the site. For each one, decide: does this page deserve to exist at full length? If yes, expand with worked examples, decision frameworks, or case-study tie-ins. If no, sunset the page (301 redirect to a relevant pillar). Resist the middle path of “pad it to 1,000 words with generic content” — that is the worst of both options.
11 Schema FAQPage schema that does not match visible body content costs rich-result eligibility AND AEO citation.
The mechanism. Google's Rich Results Test flags “structured data does not match content” when FAQPage JSON-LD contains questions or answers that do not appear verbatim in the rendered body. Beyond the obvious rich-result-eligibility hit, AI retrieval systems read the same mismatch as a trust violation: the page is making a structured claim it cannot back up with rendered text. Citation rate drops measurably even after rich-result eligibility is restored, because the model has flagged the page as low-trust.
How the mismatch happens. Most often by accident. A content team rewrites the rendered FAQ section but forgets to update the JSON-LD; a developer copies a JSON-LD template into a new page and never updates the questions to match the page's actual content; a CMS plugin generates JSON-LD automatically from page metadata that does not match what is displayed. All three are common.
How to detect it. Run the Rich Results Test on every page with FAQPage schema, weekly during content-heavy periods, monthly otherwise. The test flags mismatches explicitly. For a quick spot-check: open the page, view source, search for application/ld+json, and compare the question text in the JSON-LD against the rendered question text on the page. If they differ in any way (punctuation, word order, missing words), there is a mismatch.
How to apply. Establish a content-team rule: any FAQ rewrite must include a JSON-LD update. Treat FAQPage schema as part of the content, not as an afterthought. For sites using JSON-LD plugins, audit the plugin output against rendered content and either trust it (after verification) or replace it with hand-maintained schema in the page template.
The recovery curve. After fixing a mismatch, rich-result eligibility recovers within 1–2 crawl cycles (typically 3–7 days for active sites). AEO citation recovers more slowly; we have seen 4–8 weeks for the trust signal to fully restore based on the Search Console structured-data report tracking.
12 Multimodal Image alt text containing the brand or entity name contributes to AEO citation rate.
The mechanism. AI surfaces increasingly index multimodal content; image alt text is the lowest-cost multimodal signal you can ship. When alt text contains the brand name, the operator name, or a specific entity reference, the alt text becomes a quotable atom even when the image itself cannot be retrieved. This is most visible on Google AI Overviews (which sometimes cites image alt text in image-adjacent queries) and on Perplexity (which reads alt text as part of the page's overall semantic content).
What “containing the brand name” looks like in practice. Bad: alt="hardwood floor". Better: alt="refinished red oak hardwood floor in a Bentonville Arkansas home". Best: alt="Heritage Hardwood Floors NWA refinished red oak hardwood floor in a Bentonville Arkansas home, 2026". The best version is more verbose than typical SEO-friendly alt text, and that is the point — AEO context windows reward specificity that traditional SEO would penalize.
Variation observed. The effect is most pronounced on local-business queries with image-rich answer formats. We have seen GBP's photo carousel surface alt-text-rich entries first when the query has a strong visual component (e.g., “modern flooring options Bentonville Arkansas”). On purely textual queries, the alt-text effect is weaker but still measurable.
The accessibility-first counterargument. Long, verbose alt text degrades the screen-reader experience for visually impaired users. The right answer is to use the longest accessibility-appropriate alt text, not to maximize alt text for AEO. For decorative images, use empty alt (alt=""); for content images, write alt text that communicates the image's function for someone who cannot see it, then check whether the brand name fits naturally. Never sacrifice accessibility for AEO; in this case, the disciplines align more often than they conflict.
How to apply. Audit alt text across the site. For every image whose function is to communicate something specific to the visitor (not pure decoration), include the brand name or specific entity references in the alt text where it reads naturally. Skip images where the alt text would feel forced. The audit-and-fix typically takes 2–4 hours per 30 pages.