AI Overview Optimization for B2B Service Pages

Most B2B teams are tuning their service pages for Google's AI Overview using a file Google has publicly confirmed it ignores. We run AI outbound for 50+ B2B companies and run their buyer queries through Google every month, and the pattern has never once changed: the pages that get quoted are the pages that already ranked and already answered fast. Below is how the Overview actually picks a source, the capsule format that gets lifted, what schema does and does not do, and where this whole channel stops working.

Why Does the AI Overview Decide Your Shortlist Before Your Page Does?

An AI Overview is the answer Gemini writes above the organic results, assembled from pages Google has indexed and judged trustworthy. For a B2B buyer opening a category search, that block is often the entire first impression. They read a summary of what good looks like, see a handful of named sources, and build a shortlist without scrolling.

The coverage numbers explain why this stopped being optional. Omnibound's tracking puts AI Overviews on roughly 48% of Google queries as of March 2026, up from 34.5% in December 2025. B2B Tech triggers them on 82% of queries, behind only Healthcare at 88% and Education at 83%. Your category almost certainly has one.

The behavior change underneath is sharper. SparkToro's analysis of Similarweb clickstream data, reported by Search Engine Land, found 68.01% of US Google searches between January and April 2026 ended without a single click, up from 60.45% in 2024. Seer Interactive measured organic click-through rate falling roughly 60% on queries where an Overview appears. The buyer is being answered, not routed.

This hits B2B harder than consumer search because the first query is almost never your brand. It is a problem or a category, something like "best appointment setting service for SaaS" or "how outsourced SDR pricing works," and the Overview answers it directly. Machine Relations research found 94% of B2B buyers now use AI in some form for vendor research. Being named in that answer is the modern version of ranking on page one, except the buyer never sees the page of competitors you used to fight on.

AI Overview

Google's AI-generated answer block at the top of search results, written by Gemini from indexed pages it trusts. It summarizes the query, names sources, and often satisfies the buyer without a click, which is why being one of its cited sources now matters as much as a traditional ranking.

AI Overview optimization

The practice of structuring a page, and earning authority for its domain, so Google's AI Overview pulls from it and attributes it by name. It shares the ranking half of its work with traditional SEO, then adds extraction-friendly structure, entity clarity, freshness, and third-party consensus on top.

How Does Google Pick Which Pages the AI Overview Quotes?

The selection runs in two stages, and almost every mistake we see comes from treating it as one. Stage one is retrieval. Google pulls the pages most relevant to the query, which is overwhelmingly the pages already ranking well. Search Engine Land's analysis found 40 to 76 percent of AI Overview citations also appear in the top 10 organic results, with a first-position link carrying roughly a 53 percent chance of citation against 36.9 percent at position ten. If your service page is not ranking, it is not in the candidate pool at all, and no amount of markup changes that.

Stage two is synthesis. Among the pages it retrieved, Gemini picks the ones it can lift a clean, confident answer from. That favors content answering the query directly in the first line of a section, carrying recent timestamps, and reading as authoritative rather than promotional. Freshness is a real weight here, with 85% of AI Overview citations coming from content published in the last 3 years.

The practical read is blunt. Rank the page, then make it the easiest and most specific answer in the retrieved set. Those are two separate jobs with two separate playbooks, and the reason so many teams stall is that they do the second job on a page that never cleared the first. We break the split down further in GEO versus SEO and how to rank in AI search.

Query fan-out

The way Google breaks one search into several related sub-questions and assembles the AI Overview from pages that answer each piece. A service page covering the main question plus its obvious follow-ups can be pulled into multiple parts of a single answer, which is why one page with 8 buyer-question sections outperforms 8 thin pages.

What Is an Answer Capsule and How Do You Write One?

The answer capsule is the highest-leverage block on the page and the one almost nobody publishes. It is a short standalone paragraph sitting directly under a heading, answering that heading completely, written so an engine can quote it without editing. Neither of the two content agencies we compete with in this category ships one anywhere in their libraries, which makes it the cheapest structural win available.

Answer capsule

A 40 to 60 word standalone paragraph placed directly under a heading that answers the question in that heading completely, without depending on any sentence around it. It repeats the noun phrase of the query, carries one specific number, and avoids pronouns that point outside itself. It is the unit an AI answer lifts verbatim.

Six rules make a capsule liftable. Break any one of them and the engine has to rewrite you, and an engine that has to rewrite you usually quotes someone else instead.

1. 40 to 60 words. Short enough to drop into an answer whole, long enough to be a real answer. Under 30 words reads as a fragment. Over 80 gets truncated somewhere you did not choose.
2. Standalone. No "as we mentioned above," no "this approach," no pronoun whose antecedent lives in the previous paragraph. If the capsule cannot be copied into a blank document and still make sense, it is not a capsule.
3. Repeat the noun phrase of the query. If the question is "AI Overview optimization for B2B service pages," those words belong in the capsule. Not as keyword stuffing, as the subject of the first sentence.
4. Lead with the verdict. The answer goes in sentence one. Context, caveats, and mechanism go in sentences two and three. A capsule that builds to its point gets cut before it arrives.
5. Carry one specific number. A capsule with a sourced figure in it is dramatically more quotable than one without, because the number is the part an answer engine wants and cannot invent.
6. Match the heading exactly. The heading asks a question the buyer types. The capsule answers that question and nothing else. One heading, one question, one capsule.

The difference this makes is visible in the first line of a section. Here is the same information written three ways, from the version a brochure ships to the version an engine lifts.

Notice what the right column does not do. It does not sell, it does not hedge, and it does not warm up. It answers. The page can still sell 200 words later, and it should, but the first block under every heading belongs to the question. We use this format on every page we publish, including the ones on reverse outbound and cold email reply rate benchmarks.

Then phrase the headings themselves as the questions buyers type. "How much does outsourced appointment setting cost" matches a real query. "Our Approach" matches nothing. The fan-out latches onto question headings, which is why a page with 8 buyer-question sections can be pulled into 8 different answers while a page with 8 brand-voice section titles gets pulled into none.

What Schema Should a B2B Service Page Use for AI Search?

Schema is where the advice online splits hardest, so here is the honest version. Google's own documentation on AI features says there is no special schema.org structured data you need to add for AI Overviews or AI Mode. Meanwhile third-party studies report sites with complete markup seeing meaningfully higher citation rates. Both can be true, and the resolution matters more than either number.

What schema actually does in 2026, per Digital Applied's breakdown, is help the engine verify claims, connect entities, and judge source credibility while it writes the answer. It does not push you into the Overview. It makes you easier to trust once you are in the candidate pool, and it materially helps on Bing-powered surfaces, which is what ChatGPT Search runs on. The correlation studies are mostly picking up the fact that sites with disciplined markup are also sites with disciplined content.

So build the stack, and build it for the right reason:

• Organization, with a description that names your category in a noun phrase. If a machine compresses your description to one clause, that clause needs to still place you in the right category. Negating a category ("not a production agency") compresses badly and can hand the win to a competitor.
• Service, one per offering, with plain descriptions of what is delivered. This is the block that answers "what do they actually do" when an engine is deciding whether you fit the query.
• FAQPage, carrying the questions buyers type verbatim, with answers written as capsules. This is the single highest-value block on a service page because the questions and the answers are both extractable units.
• Person, for a named author with a real bio, credentials, and sameAs links to their profiles. Anonymous marketing copy reads as lower trust than a named practitioner, and the sameAs array is how the engine confirms the person exists.
• The optional properties. author, dateModified, sameAs, and description are the ones that give an engine the context to cite you confidently instead of skipping you. Most implementations ship the required fields and stop, which leaves the useful half on the floor.

One rule that costs nothing and gets broken constantly: your schema has to agree with your prose. A page arguing it is a podcast agency while its markup declares a cold email agency will be read by the markup. We learned that on our own site, where 302 pages carried a service description the blog posts had already outgrown. Our full markup checklist sits in the generative engine optimization checklist.

Does llms.txt Actually Do Anything in 2026?

This is the part of the playbook that has been oversold, and it is worth being direct about it because teams are spending real hours here.

Ahrefs analyzed 137,210 domains receiving traffic in May 2026. About 28% published an llms.txt file, roughly 38,000 domains. Of those, 97% saw zero requests for the file in the entire month. Nothing fetched them. Among the 3% that did get traffic, 96% of requests came from bots, and the AI retrieval bots the file was invented for accounted for 1.1% of requests. SEO audit tools were the single largest requester at 21.7%.

Google's position is equally clear. Search Engine Journal reported Google confirming the file has no current implementation in Search, and Search Engine Roundtable covered John Mueller comparing it to the long-deprecated keywords meta tag, a vendor-controlled field with an obvious spam problem. His framing was that llms.txt is not built for search at all, and is closer to a temporary token-saving convenience for AI coding tools reading developer docs.

So what do you do with that? Ship the file, because it takes 20 minutes and costs nothing, and the standard could get adopted. Then never count it as an AI visibility strategy and never let it displace a single hour of work on the page itself. We publish one at highticketaisystems.com/llms.txt and treat it as exactly what it is, a cheap option on a future that has not arrived. The longer version is in llms.txt for B2B companies.

The broader lesson repeats across this whole space. Any tactic that promises visibility without ranking, without structure, and without third-party consensus is selling a shortcut around the three things that are actually doing the work. There is no file you can upload that substitutes for being the best answer.

Which Authority Signals Get a B2B Domain Trusted?

Structure gets you retrieved and extracted. Authority decides whether the engine puts your name in the answer at all. B2B service categories are exactly where Google leans hardest on trust signals, because a wrong recommendation on a high-stakes purchase is expensive, and this is the slower half of the work that most teams skip entirely.

The single most useful finding here is where citations come from. Across industry studies, earned media accounts for 82 to 85 percent of AI citations on the general web, and for software categories specifically, AI engines cite earned media at 72.7 to 74.2 percent versus 31.8 to 45.4 percent for Google. Data-Mania's B2B SaaS benchmarks land in the same place. Answer engines trust what other people say about you far more than what you say about yourself, and by a wider margin than classic search ever did.

The three surfaces also do not share rules, which is why a single checklist tends to underperform on all of them.

Four moves build the authority half, in rough order of return:

• Publish head-to-head comparisons naming real competitors. Citera's analysis of 350,000 B2B SaaS articles found pages with clear named comparisons see a 38% lift in citation rate, climbing to 51% inside ChatGPT. Most B2B teams refuse to name competitors, which is precisely why the pages that do get quoted.
• Earn mentions on independent sites. Review profiles, guest articles, and podcast appearances feed the third-party consensus that tells an engine your brand is real. One mention on a recognized outlet outweighs another post on your own blog, and this is the mechanism behind the 82 to 85 percent earned media figure.
• Publish original data. A benchmark only you can report gives other sites a reason to cite you, which compounds your authority and hands the engine a sourced statistic with your name attached. Our own numbers on the state of AI outbound and how AI is changing sales development do that job.
• Keep your entity consistent everywhere. Your name, category, and description should read the same on your site, your LinkedIn, your review profiles, and your press. Conflicting descriptions make it harder for an engine to form a confident picture, and a confused entity gets skipped in favor of a clear one.

There is one earned-media asset most B2B companies overlook, and it is the one we build our entire business around. A recorded conversation with a named guest produces a transcript, a page, and a third-party name attached to your domain, all at once. It is the rare authority signal you can manufacture on a schedule rather than pitch for. We cover the mechanics in podcast transcripts for AI search and how to get your podcast cited by AI.

How Do You Track Whether AI Overviews Cite You?

You cannot improve what you do not watch, and the measurement gap here is enormous. Only 14% of marketers track AI citation visibility at all, while 43% name AI optimization a core 2026 strategy. That gap is the opportunity: most of your competitors are working this channel blind.

The tracking loop is simple enough to run in 30 minutes a month, and simple is what makes it survive past week three.

1. Fix a list of 10 buyer queries. The category questions, the "best service for X" questions, the pricing questions, and the comparison questions. Same 10 every month, because the point is the trend line.
2. Log three fields per query. Did an AI Overview appear, which sources did it name, and were you one of them. A spreadsheet is enough. The named-sources column is the most useful one, because it tells you exactly who the engine currently trusts in your category.
3. Check the underlying rank in Search Console. A page that fell out of the top 10 will fall out of the Overview with it, so a citation you lost is usually a ranking you lost first.
4. Repeat the queries in ChatGPT and Perplexity. Different retrieval, different answers, same 10 questions. Our approach to that is in how to track AI search visibility and how to audit your brand in ChatGPT.

Expect the trend to move slowly. Crawl, index, and re-embed all lag, so a fix shipped today shows up 4 to 8 weeks later on a corpus of any size. That lag is exactly why the monthly cadence beats checking obsessively, and why teams that judge this channel after 2 weeks always conclude it does not work. More on the attribution problem in LLM citation versus SEO traffic and why AI answers cite some companies and not others.

Where Does AI Overview Optimization Stop Working?

Here is the limit nobody selling GEO services will tell you. Every technique on this page only reaches a buyer who already typed the question. That is a real and valuable slice of the market, and it is a fraction of your addressable market on any given month.

The zero-click math makes it sharper. If 68% of searches end without a click and click-through falls 60% when an Overview appears, then winning the citation frequently means winning a mention with no visit attached. That is still worth having, because the mention shapes the shortlist. It is just presence, not traffic, and a business planned around it as a traffic channel will be disappointed by numbers that are actually performing fine.

Which is why we treat AI visibility as one half of a machine rather than the whole thing. The other half reaches the buyer who has not asked the question yet. For us that is a podcast invite sent by email: we invite a client's ideal buyers onto their own show, the recording does the qualifying work a discovery call would have done, and the client owns the edited episode. Editing is included, invites are email only, and we back it with 30 recorded conversations with your ideal buyers in 90 days or your money back.

The two halves feed each other, which is the part that surprises people. Every recorded conversation produces a transcript and a page carrying a third-party expert's name on your domain, which is the earned-media signal answer engines weight at 82 to 85 percent of citations. The outbound builds the authority that wins the Overview, and the Overview pre-qualifies the buyers the outbound reaches. Start with podcast lead generation for B2B, then how to invite guests to your B2B podcast and how to turn podcast guests into clients.

The invitation layer is where that machine actually breaks, and it breaks on infrastructure rather than strategy. An invite that lands in spam books nothing regardless of how good the show is. That means email deliverability, domain reputation, sending domains, warmup, SPF, DKIM and DMARC, list verification, inbox placement testing, multi-domain sending, and staying out of the spam folder. Then the targeting layer, defining an ICP and knowing how many invites it takes to book one recording. That stack is unglamorous and it decides everything downstream.

The Practitioner Takeaway

AI Overview optimization is not a new discipline bolted onto SEO. It is ranking, plus extraction, plus trust, in that order, and the order is not negotiable. Rank the service page first, because the Overview almost always quotes pages already in the top results. Then make it the easiest answer in that set: a 40 to 60 word capsule under every buyer-question heading, sourced numbers instead of adjectives, one comparison table, and self-contained sections.

Ship the schema stack for what it really does, which is help an engine verify who you are and what you sell, not conjure a citation out of nothing. Ship llms.txt in 20 minutes and then forget about it, because Google says Search ignores it and 97% of the files published never get read. Spend the hours you save on the thing that actually moves the number, which is earning mentions somewhere other than your own domain.

Then run your 10 buyer queries every month and watch the named-sources column. That column is the scoreboard, and it will tell you the truth long before your analytics do. Being quoted is the same job as being recommended, just done by a machine reading the web instead of a peer in a Slack channel. The companies that compound do not choose between getting found and reaching out. They earn the citation that pre-qualifies the buyer who is already searching, and they run the invite that reaches the buyer who has not started looking yet.