Research library

Google's robots.txt token for blocking AI training (Google-Extended) blocks Gemini training — but does NOT exclude a site from being used to generate AI Overviews answers. Google treats AI Overviews as a Search surface.

A Princeton study (KDD 2024) tested nine content strategies across ~10,000 queries. Adding statistics, citations, and direct quotations lifted AI citation visibility by up to ~40% — and sites that don't already rank #1 gained disproportionately more. Important qualifier, recorded openly: a later multi-domain benchmark (C-SEO Bench, §9.6) found these tactics largely failed to replicate at scale — traditional SEO outperformed them. We cite both, and our recommendations lean on fundamentals first.

Yext analyzed 6.8 million citations across 1.6 million AI responses. Gemini cited brand-owned websites 52% of the time; ChatGPT cited third-party directories (Yelp, TripAdvisor, MapQuest) 49% of the time; Perplexity leaned on niche industry directories (~24% — on unbranded, subjective queries, the one query type Yext reports that figure for). This is why Gault's per-engine actions differ. (Updated 2026-06: for over a year this entry ended "Claude has no action yet — no study covers Claude's sources, and we don't guess." Yext's 17.2M-citation follow-up (§9.17) finally covers Claude, so Claude now has an action — added the day the receipt existed, not before.)

Semrush's citation tracking documented Reddit's share of ChatGPT citations collapsing from ~60% to ~10% in about six weeks in 2025 — and Similarweb's larger 600K-citation study (§9.8) measured Reddit at ~12% of ChatGPT citations in Jan–Feb 2026 — consistent with a re-level rather than a glitch (Similarweb reports the one snapshot; joining the two series is our reading). AI visibility is volatile; that's why Gault tracks it over time instead of treating one scan as the truth. (Entry updated 2026-06: Semrush has since rewritten its study page with a larger dataset; the per-domain percentages we originally quoted from it now live in §9.8's fresher numbers.)

Updated 2026-06 and again 2026-09 (each change recorded, never silently swapped): Similarweb's 2025 measurement put ChatGPT at ~89% of AI referral visits — that figure is superseded and no longer on Similarweb's page. The current numbers measure something different: the share of visits TO the AI chatbots themselves, not referrals FROM them to websites. As of May 2026, ChatGPT holds roughly 53% of visits to generative-AI sites (Gemini ~27%, Claude ~9%). We don't have a sourced number for ChatGPT's share of referrals to business websites; an earlier '63% of B2B referrals' figure was removed because we could not find it on any page we can link. ChatGPT is still the clear #1 by usage, which is why Gault's AI Visibility Report asks it every query in your industry's pack while other engines get the primary question — but the margin is narrowing, and this rule is under standing review as Gemini grows.

Google's own documentation: local ranking comes from relevance, distance, and prominence — and "more reviews and positive ratings can help your business's local ranking." Google instructs businesses to keep info complete and accurate, keep hours up to date, respond to reviews, and add photos. Note the careful wording: photos and review responses appear under profile completeness, not ranking factors.

Google's representation guidelines: your profile name must reflect your real-world business name — adding keywords, locations, or taglines "isn't permitted, and could result in the suspension of your Business Profile." On categories: use as few as possible to describe your core business. This is why Gault flags stuffed names as a compliance risk, not a ranking tip.

Whitespark's Local Search Ranking Factors 2026 (47 experts, 187 factors): GBP signals are the largest Local Pack factor group (~32%, from the survey's group-weighting chart), 8 of the top 10 factors come directly from the profile (our count of Whitespark's list), primary category sits at the top, and the survey's own headline is that review signals rose in importance (recency ranks #11, a sustained influx of reviews #14). It's an expert-opinion survey, not a measurement study — Gault treats the trends as consensus and the exact weights as informed opinion.

BrightLocal's consumer survey (1,002 US consumers, methodology disclosed): 68% will only use a business rated 4.0+, and 31% now require 4.5+. 74% only care about reviews from the last three months. 89% expect owners to respond to reviews. And the AI crossover: AI tools jumped from 6% to 45% as a local-recommendation source in one year. These are trust-and-conversion findings — they shape what customers do, which is reason enough.

Sterling Sky's client case data (cited by Whitespark's Darren Shaw, who ranks review recency in his top 5 ranking factors for 2025): rankings dropped when new reviews stopped flowing and recovered when they resumed. Raw review count matters at a low threshold and then plateaus — but a steady, recent inflow keeps mattering. This is why Gault weights review velocity, not just total count.

A controlled test (Sterling Sky) found adding photos had no measurable Local Pack ranking impact. A 2-million-profile analysis (Localo) found top-3 listings average 250+ photos vs ~170 for positions 11–20 — real, but correlation. Google officially recommends photos as part of a complete profile. Honest synthesis: photos matter for the people viewing your profile, not for rank — and that's why Gault frames its photo checks as engagement, never ranking.

Google encourages replying to reviews but attributes the benefit to standing out, not ranking. 89% of consumers expect owners to respond (§7.4), and the average response rate across SOCi's 350,000-location index is under half — so responding consistently is a real differentiator. The widely repeated "responding boosts rankings" claim traces only to tool-vendor blogs with no methodology, so Gault treats responding as a customer-expectation check. (Entry updated 2026-06: an earlier "~80% vs ~45%" leaders-vs-average stat was removed — its only live source no longer carries it.)

Sterling Sky's category tests found that adding genuinely relevant secondary categories expands the searches a profile appears for without diluting its primary-category ranking — one documented example saw a law firm add "Employment Attorney" and improve for employment queries within days (the ~48-hour figure comes from the author's follow-up comments, not the article body — we're precise about provenance). Google's guideline says use as few categories as possible, so the rule is relevance: helpful when the category truly fits, a guideline risk when you stack marginal ones. That's why Gault's check requires relevant categories.

A controlled test (Sterling Sky) — one post a week for nine weeks across three listings, 441 tracked keywords each — found posting had no measurable impact on Local Pack rankings. Posts still help engagement and keep a profile looking active. Gault keeps its recent-post check as a freshness/engagement signal and its copy never implies a ranking effect.

Sterling Sky tested keyword-rich Q&A across multiple listings: no ranking impact. Separately, Google discontinued the Business Profile Q&A API in November 2025 with no replacement — the feature is being wound down in the AI-answers era. Gault removed Q&A from scoring accordingly.

Since late 2023, local-SEO practitioners have observed that whether a business is currently open affects Local Pack rank: Sterling Sky published before/after screenshots for a lawyer, a psychiatrist and (via a third party) restaurants, and relayed a forum report of a listing dropping from #1 to #10 overnight while closed. That is observation, not a controlled test, and the author's own wording is that Google 'appears to be' using it. Separately, Google says in writing that for broad queries, temporarily-closed businesses can rank after open ones — a statement about the closed STATUS, not about hours. Accurate hours are the precondition for the observed signal working in your favor. (Updated 2026-09: earlier wording called this 'replicated' and 'measured'; the source doesn't support that strength.)

SOCi analyzed 350,000+ locations across 2,751 multi-location brands: AI platforms recommend only 1.2% of locations on ChatGPT, 7.4% on Perplexity, 11% on Gemini — versus 35.9% getting Google 3-pack visibility. In retail, only ~45% overlap between the top-20 brands winning traditional local search and the top-20 recommended by AI. Recommended locations average ~4+ stars: ratings act as an inclusion filter. This gap is why AI visibility is its own pillar in Gault, not a footnote to the Map Pack. (These figures are reported in SOCi's 2026 Local Visibility Index, downloadable from the linked page. SOCi competes in this category — framed accordingly.)

Sterling Sky's single-variable testing (Feb 2026) found that populating the structured GBP "Services" field produces a measurable Local Pack ranking lift, with documented before/after examples across multiple industries — the effect typically appears within 24–72 hours. Sterling Sky calls it a powerful, under-utilized lever. This grounds Gault's Services check, which replaced an older description-keywords check that had no ranking evidence.

Google's title-link documentation instructs that every page carry a descriptive <title>, and lists the title element first among the sources it uses to generate the clickable title in results (og:title and the first visible H1 are also listed). Google frames titles as critical for users; it doesn't quantify a ranking weight. Gault's title and H1 checks rest on this.

Google may use the meta description for a result's snippet when it describes the page better than other on-page text — a snippet and click-through quality input, not a ranking signal. Gault keeps its meta-description check, framed honestly as a click-through (not ranking) lever.

Google's official site-experience triad: Largest Contentful Paint ≤ 2.5s, Interaction to Next Paint ≤ 200ms, Cumulative Layout Shift ≤ 0.1 (Google's wording is 'or less') — assessed at the 75th percentile of real-user field data. INP replaced FID as the responsiveness metric in March 2024. Gault's LCP, CLS, and INP checks grade against these official bands. (The thresholds live on web.dev's Core Web Vitals page; the INP announcement carries the date and the 75th-percentile rule, not the numbers.)

Google's page-experience documentation: "Core Web Vitals are used by our ranking systems" — and in the same breath, chasing perfect scores "may not be the best use of your time," because relevance wins even when page experience is sub-par. Real signal, modest weight — which is exactly how Gault weights its speed checks.

Schema.org's authoritative LocalBusiness vocabulary defines the subtype tree Gault maps each vertical to (Dentist, Restaurant, InsuranceAgency, ExerciseGym, and so on) plus standard properties — name, address, telephone, url, geo, opening hours, sameAs. It's the basis for Gault's per-vertical schema-type matching.

Google's official local-business structured-data documentation: only name and address are required, more properties mean higher quality, and — quoted directly — "Use the most specific LocalBusiness sub-type possible." This is why Gault flags schema that uses a generic type when a specific one exists for your industry.

Google's structured-data policies state plainly that a structured-data manual action means a page "loses eligibility for appearance as a rich result; it doesn't affect how the page ranks." Schema's honest value is rich-result eligibility, disambiguation, and machine readability (increasingly relevant for AI surfaces) — never a ranking lift. Gault's schema checks are worded to match.

Google confirmed HTTPS as a ranking signal in 2014 — "a very lightweight signal… carrying less weight than other signals such as high-quality content" — and it remains a page-experience criterion today. Gault's HTTPS and redirect checks rest on this and Google's canonical-selection guidance.

Google's transition to mobile-first indexing is complete: sites that work on mobile are crawled primarily with the mobile Googlebot (a small set of mobile-broken sites still gets the legacy desktop crawler). Google's current page-experience guidance lists "displays well on mobile devices" among its criteria (§8.4). Gault's mobile-viewport check is a proxy for that criterion.

Google's John Mueller, explicitly (Google SEO office-hours, February 2021): the number of words on a page is not a quality factor or ranking factor. Gault keeps a thin-content check only as an honestly-labeled heuristic — a 50-word service page rarely answers anyone's question — never as a ranking claim.

Google made page speed a mobile ranking factor in 2018 (affecting only the slowest pages, since folded into page experience). But the Lighthouse 0–100 score itself is NOT what Google ranks on — Google's page-experience doc says Core Web Vitals (real-user field data) are what its ranking systems use, and Google sets no ranking pass score (PageSpeed Insights labels 90+ 'good' and under 50 'poor' as display bands only). Gault's mobile-50 bar sits at PSI's poor/needs-improvement boundary; the desktop-70 bar maps to no Google band — both are internal heuristics, and the check copy says so.

Google uses alt text together with computer-vision algorithms and page content to understand what an image shows — Google calls it the most important attribute for giving an image metadata, and it aids accessibility, and keyword-stuffed alt text can be treated as spam. There's no official coverage threshold (Gault's 90% bar is internal calibration, and its copy says so).

Whitespark's 2026 survey rates on-page signals the most important factor group for local organic results — with "a dedicated page for each service" the #1 factor and localized content / city-keyword placement rated highly. Important honesty point: no Google document says "put your city in your title or H1" — these are expert-consensus signals plus the general mechanism that descriptive titles aid relevance. Gault labels them as consensus, not "Google says so."

Google's documentation: rel=canonical is "a hint, not a rule" — Google selects canonicals using HTTPS vs HTTP, redirects, sitemap inclusion, and rel="canonical" annotations (its doc's four listed factors; linking consistently to the canonical URL is a separate preference hint in Google's how-to doc). Near-duplicate location pages are the natural case for this. Small, well-linked sites (~500 pages or fewer) may not need a sitemap, and malformed hreflang tags are silently ignored. This grounds Gault's canonical, redirect, sitemap, and hreflang checks.

A page carrying a noindex directive is removed from Google Search results — definitional, and the strongest-evidenced check in Gault's entire suite. If a page that should rank is accidentally noindexed, nothing else about it matters until that's fixed.

The Open Graph protocol (og:title, og:type, og:image, og:url required; og:description recommended) is a social-preview standard — not a Google ranking signal. Its two honest values: clean link previews in social and messaging apps, and machine-readable page metadata that Google's title-link docs list as a title source. Gault weights its Open Graph checks modestly to match.

Google restricted FAQ rich results to government/health sites in 2023 and fully deprecated them for all sites by 2026 — FAQPage schema now produces no Google SERP feature. But FAQ-format CONTENT remains valuable: Microsoft officially recommends Q&A formats for inclusion in Copilot answers, and a business's own site is the dominant source ChatGPT cites for local questions. So Gault's FAQ check is a content check, never a (dead) schema recommendation.

Google's March 2024 core update folded the Helpful Content System into core ranking and launched spam policies targeting scaled content abuse. For multi-location brands the relevant rule is Google's separate, longer-standing Doorway abuse policy: "multiple domain names or pages targeted at specific regions or cities that funnel users to one page" and "substantially similar pages" — so templated, near-identical city or location pages are a spam-policy risk, and location pages need genuinely differentiated local content. This is why Gault's local-content checks reward substance over boilerplate. (Updated 2026-09: the March-2024 scaled-content policy's own examples don't mention location pages; the doorway policy does.)

A dedicated page per service is the #1 local-organic ranking factor in the 2026 expert survey and #2 on its AI-visibility list, and AI engines cite deep service pages far more than homepages. Across the insurance corpus — independent practitioner blogs plus a carrier blog (Nationwide, which recommends a pillar page per protection area: home, auto, commercial, farm and specialty) — the consensus is that each product line warrants its own substantive page, not one generic services page. Live competitor inspection agrees: agencies ranking for "best auto insurance agent Houston" carry 14–18 differentiated per-product pages, while ~200-word stub pages with no product pages don't rank. The hard caveat, from Google itself: pages must be genuinely differentiated — near-identical templated product/city pages are the doorway-page pattern Google penalizes (§8.18). So Gault's servicePageCoverage check looks for at least 3 dedicated, differentiated product pages (the more the better) and never rewards near-duplicates; city-on-page is treated as a bonus, not a requirement, to avoid steering toward doorway pages. The 3-page minimum is Gault's internal calibration, and there's no controlled test isolating website pages, so the copy claims no proven ranking lift (tier B+).

Google's official AI-optimization guide (May 2026) says appearing in its AI features takes no special files, no llms.txt, no AI-specific markup — "structured data isn't required for generative AI search." Solid SEO fundamentals and Business Profile data are what feed Google's AI answers. Note the engine split: Google downplays schema for its AI surfaces while Microsoft recommends it for Copilot (§9.3) — Gault's recommendations name engines for exactly this reason.

Microsoft's official guidance for inclusion in Copilot / Bing AI answers recommends JSON-LD schema markup, clear heading structure, Q&A formats, lists and tables, and self-contained sentences — and warns against content hidden in tabs, PDFs, or images. Copilot is powered by Bing's index, so traditional Bing SEO and Bing Places listings drive AI-answer visibility on Microsoft surfaces.

Each engine now documents which crawler controls SEARCH visibility separately from training: OpenAI says blocking OAI-SearchBot means your site "will not be shown in ChatGPT search answers"; Anthropic's Claude-SearchBot and Perplexity's PerplexityBot control the same for their engines. A robots.txt that blocks these makes a business structurally invisible to AI search regardless of everything else.

C-SEO Bench — the first multi-domain, multi-actor benchmark of conversational-SEO methods, including the Princeton GEO tactics (§2.1) — found most content-rewrite tactics "largely ineffective" and sometimes negative, while traditional SEO (being the page that gets retrieved) was significantly more effective. Gains also decay as more sites adopt the same tactics. We publish the study that cuts against our own recommendation surface, because that's what a receipt trail is for.

A large-scale academic comparison of AI search engines vs Google found a "systematic and overwhelming bias towards earned media" — third-party, authoritative sources — over brand-owned and social content, plus significant engine-to-engine divergence. Independent academic convergence with the per-engine citation studies (§3.1): being written about beats writing about yourself, for AI visibility. Scope note: the earned-media bias was measured on consumer-product ranking queries (phones, cars, software); the paper's local-services experiment measured only how little AI engines overlap with Google, not source types — so applying the finding to a local business is an extrapolation we make, and label.

Similarweb analyzed ~600,000 citation events (Jan–Feb 2026). ChatGPT's top cited domains: Wikipedia 13.15%, Reddit 11.97%. The two engines measured (ChatGPT and Google AI Mode) cite many of the same domains but weight them very differently, engines rarely cite homepages (deep pages dominate), and intent shifts sourcing. This is the current grounding for Gault's Wikipedia and Reddit external signals — presence on the domains AI engines actually cite is corroboration AI tools can use.

Ahrefs tracked 1.9M US queries (June 2026 snapshot): Grok's top cited domains are Reddit (16.3%), YouTube (15.1%), and Facebook (13.9%) — heavily social/UGC. Surprisingly, X itself is only ~1.4% of citations. Grok isn't a Gault scan engine yet; if it's added, social presence carries unusual weight there. Note: the source page auto-updates monthly, so its live figures will move past this snapshot.

BrightLocal manually analyzed 800 local-intent ChatGPT queries: "business websites" — the study's catch-all for any site that isn't a directory or a mention page — made up 58% of cited local sources, brand mentions in articles and lists 27%, directories 15% — and within directories, Yelp was absent. That openly disagrees with the broader-query Yext data (§3.1), where ChatGPT leaned on Yelp-style directories heavily. Both stay cited; the honest takeaway is that your own site's content carries the most weight for local AI answers, and directory presence is corroboration, not a promised citation.

Yext Research analyzed 17.2 million AI citations (Q4 2025, location-level, methodology and limitations disclosed). Gemini favors brand-owned websites most; ChatGPT's source mix varies by industry more than any other engine; Perplexity is the most stable; and the standout: Claude cites user-generated content — reviews and other platforms where the business participates but doesn't control the content — at 2–4x the rate of other models in every sector studied (in food & beverage, ~10x Gemini). Insurance is not one of the industries in the report. The study frames the mechanism as correlation, and so do we. Also: listings make up 54.5% of distinct citation sources, while first-party pages get re-cited most per URL — breadth and depth both matter. This study is why Claude now has its own action in Gault's per-engine recommendations: grow review volume and respond to reviews.

Withdrawn 2026-09-05, the change recorded rather than silently removed. This entry used to say the cited paper confirmed that businesses mentioned earlier in an AI answer get noticed more. It doesn't: the paper studies how the ORDER OF CANDIDATES IN THE INPUT PROMPT changes an LLM's movie and book recommendations, with no businesses, no readers, and no measurement of what people notice in an answer. Gault still records each business's position in every AI answer — as a design choice we find useful, not as a research-backed claim about customer behaviour. We know of no published study linking position in an AI answer to real customer action (it stays on our open-questions list).