One playbook for showing up everywhere people search.

Every claim in this document carries a source in the appendix. Where the evidence is one study, one vendor, or contested, it is labeled. Statistics marked Data come from published studies or official platform documentation. Mixed means sources conflict or a single source stands alone. Opinion means practitioner consensus without hard numbers.

The answer replaced the list. Being the source is the new ranking.

On May 19, 2026, Google made AI Mode the default search experience worldwide, running on Gemini 3.5 Flash. The search box was redesigned for the first time in 25 years. It now takes text, images, files, videos, and open Chrome tabs as input.

What shipped at I/O 2026

• AI Mode as the default. The classic ten blue links are now a secondary tab. AI Mode answers with a synthesized response plus cited links. Data
• Query fan-out. AI Mode breaks one query into many hidden sub-queries and searches them all at once. "How to fix a lawn full of weeds" becomes searches for herbicides, chemical-free removal, and prevention. Gemini 3.5 Flash runs more of these sub-searches than earlier models, which means more pages get a shot at being cited. Data
• Search Agents. Persistent background agents monitor the web around the clock for a user's standing question (a house hunt, a price drop) and push synthesized updates. Rolling out to paid tiers through summer 2026. Data
• Agentic checkout. Universal Cart and the new commerce protocols let agents complete purchases across merchants. Mostly relevant to retail today, but it signals where local service booking is headed. Data
• Gen AI reports in Search Console. Since June 3, 2026, Search Console reports impressions inside AI Overviews and AI Mode separately. You can finally see AI visibility in Google's own tooling. Rollout started with a subset of sites. Data

The May 2026 core update told you who wins

The core update that rolled out May 21 to June 2 was one of the most volatile on record. Nearly 80 percent of top results shifted somewhere. The pattern in the winners and losers is the whole strategy in miniature:

Google said it plainly: there is no separate AI strategy.

From Google's own optimization guide for generative AI, published around I/O 2026: "Our generative AI features on Google Search are rooted in our core Search ranking and quality systems. Optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

That is the official position, and the independent citation studies mostly back it up. What earns AI citations is the same thing that earns rankings, held to a higher bar. Three things decide whether you get cited:

1. Foundational SEO. Crawlable, indexable, fast, well-structured pages. If the crawler cannot read it, no AI can cite it. Google's index feeds Gemini and AI Mode. Bing's index feeds ChatGPT and Copilot. You need both.
2. Non-commodity content. Google's guide names this directly: content with unique perspective, expert insight, original data, or first-hand experience. The test is simple. If the AI can already write your article without you, it has no reason to cite you. If your article contains photos, numbers, and outcomes only you have, citing you is the AI's only option.
3. Corroboration across the web. AI systems cross-check claims against multiple independent sources before citing a brand. Consistent facts about you on your site, in directories, in reviews, in press, and on community platforms is what makes an AI confident enough to recommend you. Third-party mentions are the strongest single lever measured: distributing the same content through independent publications lifted AI citations by a median of 239 percent, and brands are 6.5 times more likely to be cited from a third-party page than their own. Data

What Google explicitly warns against

• Special AI files and markup hacks. Google's guide says not to bother with llms.txt or similar. The independent data agrees: across 300,000 domains, llms.txt showed no correlation with AI citations, and in an Ahrefs sample of 137,000 sites, 97 percent of llms.txt files got zero bot traffic. Data
• Artificially chunking content for AI. Fragmenting pages into robot-bait passages hurts readers and does not help extraction.
• Inauthentic mentions. Buying or faking brand mentions to game the corroboration signal is treated like link schemes. Google clarified in May 2026 that manipulating AI Overviews and AI Mode is handled like manipulating search.
• Scaled AI content. Publishing floods of low-value AI articles triggered 50 to 80 percent traffic losses for violators under the scaled content abuse policy in the March 2026 update. AI-written content is not penalized as such; automation aimed at manipulating rankings is. Data

Fewer posts, deeper posts, built around what AI cannot know.

The blog is still the main engine for topical authority and AI citations. What changed is the bar: commodity posts are now worthless, structure matters more than length, and freshness matters more in AI search than it ever did in classic SEO.

Picking topics

• Build clusters, not one-offs. A pillar page (3,000 to 6,000 words) plus 8 to 15 satellite posts, linked in both directions. Sites organized this way consistently outperform keyword-by-keyword publishing, though the specific growth figures come from agency case data, not controlled studies. Expect 12 months of sustained publishing before authority compounds. Mixed
• Plan for query fan-out. AI Mode answers a query by asking its own sub-questions. Map them: for "how often should I service my furnace," the sub-questions are cost, DIY checks, timing, and failure signs. Each sub-question is a satellite post. Cover the set and you can be cited several times in one answer.
• Target what AI answers poorly. Real project outcomes, your own cost data, seasonal patterns from your own schedule, photos of actual work. An AI cannot produce these without a source. Generic "what is X" posts are the opposite: the AI answers those from memory.
• Every post links to at least one service or conversion page. That link is how topical authority turns into revenue pages that rank.

Structuring a post

AI systems extract passages, not pages. Structure decides whether your best answer is liftable. Pages with a clean H1 to H2 to H3 hierarchy get cited at roughly 3 times the rate of flat pages, and data presented in a real HTML table gets extracted about 81 percent of the time versus 23 percent for the same data buried in prose. Data

1. Title under 60 characters, phrased the way a person asks the question.
2. Key takeaways box right after the H1. Three to five bullets. This is the biggest structural change from pre-2026 blogging; both scanners and AI grab it first.
3. Table of contents for anything over 1,000 words.
4. Question-formatted H2s, each opening with a direct 40-to-60-word answer. State the answer with numbers first, then expand below it. Objective, declarative sentences. Cut "we believe" and "in our opinion."
5. Real tables for data, at least one original image per post, descriptive filenames and alt text (section 8 covers images).
6. FAQ block: 3 to 7 questions, answers under 60 words each. Note: FAQ rich results are gone from general sites (section 7), but the Q-and-A format still feeds AI extraction.
7. Visible published date and updated date. Update the modified date only when you actually change the content. Fake refreshes read as manipulation.
8. Author byline with a real bio. Name, credentials, photo, and a link to an author page carrying Person schema with sameAs links to LinkedIn and other external profiles. Google verifies authors through what the wider web says about them, not what the bio claims.
9. Conclusion with a short recap and two or three internal links to the cluster and a service page.

How often to publish

The best frequency data available is Orbit Media's annual blogger survey. It shows two winning strategies, not one: 39 percent of bloggers publishing 2,000-plus-word posts report strong results, and 37 percent of those publishing multiple times weekly do. The people getting results committed to depth or to frequency. The average post is 1,333 words, and only 9 percent of bloggers publish at the 2,000-word depth, which makes depth the less crowded lane. Data

Word count itself is not a ranking factor. Passage ranking lets a tight 40-to-100-word answer win on its own. Write what the topic needs and stop.

Refreshing and pruning

• Freshness is a first-order factor in AI search. AI-cited URLs are about 26 percent fresher than the top ten organic results, and roughly half of Perplexity's citations point to content under 13 weeks old. One analysis puts the half-life of a page's citation potential at about a year, though that figure is a single unreplicated study. Mixed
• Refresh first: pages ranking 11 to 30 (the cheapest wins), pages with impressions but weak clicks, and top performers with aging numbers or dates.
• Prune: pages with no links, no rankings, no traffic, and nothing unique to add. One documented pruning case saw clicks up 30 percent within weeks, but that is one case, not a law. Mixed
• Audit on a 6-to-12-month cycle, and put your top 20 percent of pages on a faster refresh loop given the 13-week freshness pattern in AI citations.

E-E-A-T in practice

• Experience is the amplified signal. The March 2026 core update rewarded content demonstrating first-hand experience: specific details, original outcomes, verifiable credentials. Comprehensive but impersonal pages lost ground to documented ones.
• The author is an entity. Person schema with a sameAs chain to external profiles is the mechanical core. The richer and more verifiable the chain, the more weight the byline carries.
• Do not publish: summaries of what already ranks, thin location doorway pages, LLM rewrites of other people's content, and listicles identical to every competitor's. All of it fails the commodity test.

Timing for seasonal businesses

Publish seasonal content 4 to 6 weeks before demand peaks, not during them. In home services, air conditioning repair searches jump around 250 percent in July and furnace repair around 130 percent in January, so the content that wins July was published in May. The same logic applies to any seasonal vertical. Data

The profile is the local index every AI reads first.

Local AI answers are grounded in the same data as the map pack. Gemini reads Google Maps directly, which is why its local recommendations test at near-perfect accuracy while ChatGPT and Perplexity land around 68 percent. Whatever the platform, the profile and the reviews are the raw material.

Where the ranking weight actually is

Reviews: the thresholds moved

• The acceptable rating tightened hard in one year. 31 percent of consumers now use only businesses rated 4.5 or higher, up from 17 percent in 2025. 68 percent require at least 4.0, up from 55 percent. A 4.3 average is quietly filtering out a third of the market. Data
• Reply to every review within 24 to 48 hours. No confirmed ranking effect, but reply speed and content are read by customers and by AI systems summarizing your reputation.
• Never gate reviews. Asking everyone is allowed; screening for happy customers first violates policy.
• Review text matters. Customers mentioning the service and city in review text may reinforce relevance, but the evidence is thin. Do not script reviews. Mixed

What changed in the product

• Q&A is being replaced. The Q&A API retired in November 2025 and a Gemini-powered Ask Maps rolled out in March 2026. Keep legacy Q&A tidy where it still shows, but stop investing in it.
• LSA reviews merged into GBP. A Business Profile is now required to run Local Services Ads, and the profile is the source of truth for those reviews. Missed calls lower LSA rank, so response time is now an ads lever too.
• Verification is stricter. Video verification is standard, and locksmith and garage door categories in the US and Canada require advanced verification.

Beyond Google: the other two profiles that feed AI

Apple Business Connect feeds Apple Maps, Safari, and Siri, and Siri is moving onto a Gemini-based answer engine in 2026. Bing Places feeds Bing, Copilot, and Alexa, and Bing's index is what ChatGPT search retrieves from. Claim and complete all three profiles, then keep name, address, phone, and hours identical everywhere. Citations remain a real local factor, and inconsistent data is the fastest way to make an AI drop you: the platforms cross-check GBP, Yelp, BBB, and your site before recommending anyone.

Landing pages behind the profile

• Up to 3 locations: point the profile at the homepage.
• 4 or more locations: each profile links its own location page. Multi-location profiles all pointing at the homepage showed organic demotion in 2026 testing. Mixed
• Service-area businesses: profile links a service-area page, which links city-level service pages.
• Match categories to pages. The primary category should match the linked page's content, and service pages should mirror the secondary categories.

Operating cadence

If the raw HTML does not say it, the AI never saw it.

The single most important technical fact of 2026: the major AI crawlers do not run JavaScript. GPTBot, ClaudeBot, and PerplexityBot fetch your HTML once and move on. Anything rendered client-side after load is invisible to them.

Rendering

• Server-render or statically generate every page that matters. Prices, service descriptions, FAQs, and schema must be present in the initial HTML response. Single-page apps need SSR, SSG, or a prerendering layer.
• Speed correlates with citations. Pages with first paint under 0.4 seconds averaged 6.7 ChatGPT citations against 2.1 for slower pages in 2026 measurements. Fast raw HTML wins twice: crawl completeness and user experience. Mixed
• Content order matters. Keep the DOM in reading order, one H1, logical H2 to H3 nesting, real semantic landmarks, and real HTML tables. This is the same structure work from section 3, enforced at the template level.

robots.txt for AI crawlers

Each AI vendor now runs separate bots for training, for search indexing, and for live fetches when a user clicks. For a business that wants AI visibility, the recommended posture is to allow all three classes. Search bots get you into AI answers, user bots let citations resolve, and training bots put you in the model's memory, which is the only path into corpus-first assistants like Claude's default mode. Blocking training bots is a legitimate intellectual-property stance for publishers who sell content; for a service business trying to be recommended, it works against you.

• Check your CDN before you check your file. Cloudflare has blocked AI crawlers by default for new zones since July 2025 and moved to a pay-per-citation model in July 2026. A perfect robots.txt means nothing if the CDN is refusing the bots upstream. Verify with server logs.
• AI crawl volume is real and growing: up about 43 percent year over year into Q1 2026. Expect these bots in your logs and budget server capacity accordingly. Data
• Known non-compliant bots (Bytespider, some stealth crawlers) ignore robots.txt entirely. Blocking those is a firewall job, not a robots.txt job.

Core Web Vitals

Judged at the 75th percentile of real user data (CrUX). Performance weight was strengthened in the March 2026 update, but it remains a tiebreaker between comparable pages, not a substitute for content. Measure with CrUX first; Lighthouse is a lab check, not the verdict.

Sitemaps and indexing

• XML sitemap limits: 50,000 URLs or 50MB per file, then use a sitemap index. Never include noindexed pages, redirects, 404s, or non-canonical variants.
• lastmod must be honest. Google uses it when it is trustworthy and learns to ignore it when your CMS stamps every save. priority and changefreq are ignored; stop tuning them.
• IndexNow covers the Bing side. Bing, Yandex, and others accept instant push notification of new URLs; Google does not. Since Bing's index feeds ChatGPT and Copilot, IndexNow is now an AI visibility tool, not a nice-to-have.
• Canonicals are suggestions. Google overrides them when signals disagree. Keep them self-referencing and single-hop, never chained, and remember noindex beats canonical when both are present.
• 301s, never chains. Redirect old URLs in one hop. Every URL change without a redirect burns accumulated authority.

Architecture

• Three clicks from the homepage to every page that matters. Deeper pages get crawled less and rank worse.
• Kill orphans. Roughly a quarter of pages on typical sites have zero internal links. Every published page needs inbound links from its cluster with descriptive anchor text.
• Breadcrumbs with markup give crawlers a hierarchy mesh, but they do not replace in-content links for authority flow.

Boring URLs, decided once, never changed.

URLs are the cheapest part of SEO to get right and the most expensive to fix later. Set the rules once at the template level and treat every URL as permanent.

• Lowercase, hyphens only. Google reads hyphens as word separators and underscores as word joiners. water-heater-repair is three words; water_heater_repair is one.
• Keep the whole URL under about 60 characters, with a slug of 3 to 6 meaningful words. Drop filler words unless removing them changes the meaning.
• Say what the page is. /services/tankless-water-heater-installation beats /services/page-14 for users, crawlers, and the AI systems that quote URLs in answers.
• No dates in evergreen slugs. A year in the URL forces an annual redirect or makes the page look stale. Use dates only for genuinely time-bound content.
• Blog in a subfolder, not a subdomain. example.com/blog/ concentrates authority on one domain; blog.example.com is treated as a separate site.
• Pick one trailing-slash format and 301 the other. Same for www versus bare domain.
• Strip parameters from indexable URLs. Tracking junk creates duplicate variants that split signals and waste crawl budget.
• Never change a URL without a single-hop 301. And never redesign a site without a complete old-to-new redirect map. This is the number one way sites destroy themselves during a rebuild.

Schema builds the entity, not the citation.

An honest reading of the 2026 evidence: schema markup does not directly cause AI citations. Ahrefs tracked 1,885 pages that added JSON-LD and citation rates barely moved. At retrieval time, the AI systems read visible HTML and ignore the markup. Schema still earns its place three other ways: rich results, the map pack, and keeping Google's Knowledge Graph right about who you are.

Implementation rules

• JSON-LD in the head, one connected graph per page rather than disconnected blocks.
• Use @id to link entities. Define the business once, then reference that @id from every Service, Article, and location page. One entity, many references.
• sameAs is the verification chain. Point the Organization at your Google profile, social profiles, and Wikipedia or Wikidata if you have them. Point each author Person at LinkedIn and certification pages. This chain is how Google confirms you are who the page claims.
• Markup must match the visible page. Claiming 24/7 hours in schema while the profile says 9 to 5, or marking up reviews that are not on the page, is spam territory and creates trust conflicts between your own data sources.
• Validate everything with the Rich Results Test and the Schema.org validator before shipping.

Copy-paste reference: the four blocks that matter

Keep FAQPage markup only if it mirrors a real on-page FAQ. The visible question-and-answer text is what AI systems extract; the markup is bookkeeping.

Original photos are proof. Alt text is how machines read them.

The search box takes images now, multimodal AI reads pages with vision, and original photography is one of the clearest experience signals you can publish. Stock photos prove nothing; your own photos prove the work happened.

Alt text: the decision rule

1. Decorative or redundant to the surrounding text? Empty alt: alt="". Screen readers and AI skip it, which is the correct outcome.
2. Functional (button, icon, linked image)? Describe the function or destination: "Submit contact form," not "arrow icon."
3. Carries content? Describe what the image shows in one sentence, under roughly 125 characters, with natural search terms where they honestly fit: "Technician replacing a burst copper pipe under a kitchen sink."
4. Never start with "image of" or "photo of." The tag already says it is an image. Cut to the content.

Alt text is first an accessibility requirement and second an indexing signal. It is weighted heavily for image search and multimodal retrieval in 2026, and it is the text an AI quotes when it describes your photos. Same goes for filenames: tankless-water-heater-install-austin.webp carries signal, IMG_4032.jpg carries none.

Formats and performance

• Serve WebP as the baseline, AVIF where you can. WebP saves 25 to 35 percent over JPEG at 96 percent browser support; AVIF saves another 25 to 50 percent at 95 percent support. Use a <picture> element with fallbacks.
• The hero image is the LCP element on about 38 percent of pages. Preload it with fetchpriority="high" and never lazy-load it. Google's own tests cut LCP from 2.6s to 1.9s with this one change. Data
• Lazy-load everything below the fold with native loading="lazy".
• Always set explicit width and height so the layout never shifts while images load.
• Use srcset and sizes. Only about a third of sites do; responsive images are still an easy edge.

Metadata: what counts and what is myth

• IPTC copyright and creator fields are read by Google and shown as credit lines in image results. Fill them; completeness correlates with better image rankings. Data
• EXIF camera settings do nothing. Google has said so directly.
• GPS geotagging is a myth twice over. Google strips the metadata on Business Profile uploads, and no ranking effect has ever been shown on the web side. It ranked as the number one local SEO myth in the 2026 practitioner survey. Data
• Image sitemaps and ImageObject markup help discovery for images not linked in plain HTML. Useful for galleries; optional elsewhere.
• Original beats stock in practice, not by decree. There is no official ranking bonus for original photography, but originals carry the metadata, context, and experience signals that stock never can, and they are the only images that support an E-E-A-T claim.

Six answer engines, six different doors in.

Every platform runs its own retrieval pipeline with its own biases. The foundations are shared, but the last mile differs. ChatGPT's share of AI referral traffic fell from 87 to about 63 percent in three months of 2026, so optimizing for one engine is already a dated strategy.

The cross-platform levers, ranked by measured effect

1. Earned media. The strongest measured lever by a wide margin: distributing the same story through independent publications lifted AI citations 239 percent at the median, brands are 6.5 times more likely to be cited from third-party pages, and 95 percent of AI-cited links are unpaid sources. Branded mentions across the web correlate with AI visibility about 3 times more strongly than backlinks do. Data
2. Listicle placement. "Best X in Y" roundups are the single most-cited content format, around 22 percent of all citations in 2026 studies. Getting into other people's roundups usually beats writing your own.
3. Freshness. AI-cited pages run measurably newer than organic results. A 13-week refresh loop on cornerstone content is a working rule of thumb.
4. Entity consistency. Identical brand facts everywhere: site, profiles, directories, reviews, press. Wikipedia and Wikidata entries are the deepest version of this for brands notable enough to sustain one; Wikipedia alone is ChatGPT's single most-cited source.
5. Extractable structure. The section 3 rules: question H2s with direct 40-to-60-word answers, real tables, quantified claims with named sources.

New scoreboard: impressions, citations, and AI referrals.

The measurement stack caught up in 2026. You can now see AI visibility in first-party tools instead of guessing from traffic anomalies.

The three first-party reports

• Search Console Gen AI performance reports (June 2026): impressions inside AI Overviews and AI Mode, by page and country. This is the ground truth for Google-side AI visibility.
• Bing Webmaster Tools AI Performance (early 2026): citation tracking for Copilot and Bing's AI answers. The only first-party window into the Bing-fed engines.
• GA4's AI Assistant channel group (May 2026): auto-classifies traffic from ChatGPT, Gemini, Claude, Perplexity, and 20-plus assistants. One big caveat: 35 to 70 percent of AI-referred sessions arrive with no referrer and land in Direct, so treat the channel as a floor, not a total. Data

What to track monthly

• AI impressions (Search Console Gen AI report) alongside classic impressions and clicks.
• Citation rate on a fixed prompt set: the same 10 to 20 buyer questions asked monthly across ChatGPT, Perplexity, Gemini, and Copilot, logging who gets named. Dedicated trackers (Profound, Otterly, and similar) automate this if manual checking gets old.
• AI referral sessions and their conversion rate. Small volume (about 1 percent of visits in early 2026) but converting 4.4 times better than organic. Watch the trend, not the absolute number; projections put AI referrals at 20 to 28 percent of referral traffic by year-end, though that is a projection, not a measurement. Mixed
• Review velocity and rating against the 4.5 threshold from section 4.
• Core Web Vitals at the 75th percentile from CrUX, not lab runs.
• AI crawler hits in server logs. If GPTBot and PerplexityBot stop showing up, something upstream broke; you want to notice within days, not quarters.

Citation share is volatile: ChatGPT's reliance on Reddit reportedly collapsed from around 60 percent to 10 percent of citations in six weeks in late 2025. Platform-specific tactics decay fast. The foundations in sections 2 through 8 are what hold.

The whole playbook, in execution order.

Ninety days, three phases. Foundation before content, content before authority. Skipping ahead wastes the later work.

The standing rules

• Every page passes the commodity test before it ships: what does this contain that an AI could not write without us?
• Every claim gets a number, every number gets a source. Vague attribution is a citation disqualifier for AI systems and readers alike.
• Every image is original where proof matters, has honest alt text, and ships as WebP or AVIF with dimensions set.
• Nothing structural changes without a redirect plan.
• Freshness is scheduled, not aspirational: cornerstone content on a 13-week loop, full audit every 6 to 12 months.
• Verify the fundamentals quarterly: AI facts in this document decay fast; re-check anything load-bearing before betting on it.

Where the numbers come from.

Compiled July 3, 2026 from six parallel research sweeps. Primary sources (platform documentation and official announcements) outrank trade press, which outranks vendor studies. Single-source claims are labeled Mixed in the body text.

Google official

• Google Search I/O 2026 updates: AI Mode, Search Agents (May 2026)
• Google's guide to optimizing for generative AI features
• Search Central: new resource for generative AI optimization (May 2026)
• Search Console Gen AI performance reports (June 3, 2026)
• Gemini 3.5 announcement (May 23, 2026)
• Google guidance on AI-generated content
• Google URL structure guidelines
• LocalBusiness structured data reference
• FAQPage documentation (rich result restriction)
• Google image SEO best practices

Platform documentation

• OpenAI crawler documentation (GPTBot, OAI-SearchBot, ChatGPT-User)
• Bing Webmaster Tools AI Performance (Feb 2026)
• Cloudflare AI crawler controls and Pay Per Crawl
• W3C alt text decision tree

Studies and data

• AI Platform Citation Source Index 2026 (680M citations)
• Ahrefs: 1,885 pages adding schema, AI citations barely moved (May 2026)
• SEJ: llms.txt shows no effect across 300k domains
• Earned media and AI citations: the evidence base (Stacker/Scrunch, Superlines, Muck Rack, Ahrefs)
• Content freshness in AI search (Ahrefs 17M citations)
• Orbit Media annual blogger survey
• Zero-click searches reach 69% (2026 study)
• AI Overviews cut clicks 42% by Q4 2025
• Lantern: AI referral traffic and 4.4x conversion
• GA4 AI Assistant channel and the referrer gap

Local search

• Whitespark Local Search Ranking Factors 2026
• BrightLocal Local Consumer Review Survey 2026
• Whitespark: geotagging photos is a myth
• AI local visibility report 2026
• How different AI engines generate and cite answers

Core updates and trade press

• May 2026 core update rollout
• May 2026 core update winners and losers
• AI crawlers explained (2026)
• llms.txt status and Google statements
• E-E-A-T and the March 2026 experience shift
• Seasonal search shifts in home services