The AEO, GEO & SEO glossary

It is already in mainstream products. Chrome’s auto browse, launched in preview for US AI Pro and Ultra subscribers in January 2026, can fill in forms, get quotes and schedule appointments. Agents read a page through screenshots, the HTML and the accessibility tree, so what helps them is mostly good web practice:

• real <button>, <a> and <label> elements rather than clickable <div>s;
• key facts such as prices and opening hours as text in the HTML, and no hover-only controls;
• a stable layout that doesn’t shift while the agent acts.

Lighthouse added an agentic browsing category in May 2026 that checks some of this. A Chennai restaurant whose table booking is a labelled HTML form is far easier for an agent to complete than one that only takes bookings through an image or a chat widget. WebMCP goes a step further.

Blocking one agent does not block the others, so it helps to know the names each company publishes:

• Training: GPTBot (OpenAI) and ClaudeBot (Anthropic). Google has no separate training crawler; it uses the Google-Extended token explained under robots.txt.
• Search index: OAI-SearchBot (ChatGPT search), Claude-SearchBot and PerplexityBot. OpenAI says sites that block OAI-SearchBot will not be shown in ChatGPT search answers.
• User-triggered fetch: ChatGPT-User, Claude-User and Perplexity-User load a page because someone asked. OpenAI says robots.txt rules “may not apply” to these visits, and Perplexity says its user agent “generally ignores” them.

Many businesses allow the search and user agents and decide on training separately. Our free checker shows which of these your robots.txt currently lets in.

The basic method: list the questions real customers ask, run each one on the engines that matter to you, and record which brands are mentioned, cited or recommended. Your share is your appearances divided by all brand appearances across the set.

Answers vary by engine and by day. Orbit Media’s September 2026 study of ChatGPT, Claude, Gemini and Perplexity found all four cited the same domain for the same question only 1.7% of the time, and it advises judging trends on three or more weeks of data. So track each engine separately, repeat runs, and read the direction rather than one day’s figure. Bing’s citation share metric is an official version for Microsoft’s AI surfaces. See AI visibility tracking.

The test: someone who reads only the first sentence under a heading should already have their answer. If the heading asks “How long does GST registration take?”, that sentence gives the typical timeframe and what changes it, not a history of GST; steps, exceptions, a source and an example follow.

A few habits make it work:

• use the question people actually ask as the heading;
• keep the answer to two or three plain sentences that make sense out of context;
• put numbers, prices and names in the text, not only in images or PDFs.

It is good writing, not a trick. Google’s AI guide says to write for your human audience with clear headings and sections, and that you don’t need to write in a special way for generative AI search. Our AEO guide has worked examples.

Unlike Google’s report, it measures citations rather than impressions. Metrics include total citations, average cited pages, grounding queries (the key phrases the AI used when retrieving content it went on to cite) and page-level citation activity over time.

A June 2026 update grouped grounding queries into intents and topics, added period comparison, and introduced citation share: your citations as a percentage of all citations shown across all sites for the same grounding query. That makes it the closest thing to an official AI share of voice figure, though only for Microsoft’s surfaces.

To see it, verify your site in Bing Webmaster Tools. Read the numbers as trends over weeks rather than day-to-day verdicts.

Ahrefs’ study of 75,000 brands (updated August 2026) found YouTube mentions correlated at about 0.74 with visibility in ChatGPT, AI Mode and AI Overviews, and branded web mentions at 0.66–0.71, while backlinks correlated only weakly. Ahrefs stresses that correlation isn’t causation. Semrush found LinkedIn was the second most-cited domain in ChatGPT Search, AI Mode and Perplexity answers in its dataset.

Earn mentions where your customers already look: trade press, association websites, YouTube explainers, LinkedIn posts and local news. Don’t publish self-ranked “best” lists. In Lily Ray’s June 2026 analysis of AI Overviews, when a brand’s own listicle was cited, that brand was left out of the recommendation 69% of the time. See entity SEO.

Duplicates are normal: http and https, www and non-www, tracking parameters such as ?utm_source=, and filter or sort options on a shop all create extra URLs for the same content. Google crawls the canonical most regularly and the duplicates less often.

To steer the choice, send consistent signals:

• 301 redirects and rel="canonical" are strong signals; listing a URL in your sitemap is a weak one.
• Link internally to the canonical version only.
• Don’t use robots.txt or noindex to pick a canonical; Google advises against both.

If URL Inspection in Search Console shows a different Google-selected canonical from the one you declared, look for mixed signals, such as internal links, sitemaps and redirects pointing at different versions.

Cloudflare’s version, launched in September 2025, is a robots.txt line such as Content-Signal: search=yes, ai-input=yes, ai-train=no. search covers classic results with links and short excerpts, ai-input covers grounding and AI-generated answers, and ai-train covers training or fine-tuning models.

Treat it as a public statement, not a control. Cloudflare itself says some companies might simply ignore the signals, and Google’s robots.txt parser skips lines it does not support. The IETF’s AI Preferences working group is drafting a shared vocabulary (AI training, AI use and search), still an Internet-Draft in September 2026.

Adding the line costs nothing and records your position. To actually keep a bot out, use user-agent rules in robots.txt or block it at your firewall.

The “good” thresholds, measured at the 75th percentile of page loads, with mobile and desktop assessed separately:

• LCP: 2.5 seconds or less;
• INP: 200 milliseconds or less (INP replaced First Input Delay in March 2024);
• CLS: 0.1 or less.

Google says Core Web Vitals are used by its ranking systems, but good scores don’t guarantee top rankings, and it will still show the most relevant page even when its experience is sub-par. Treat them as a baseline for users rather than a ranking lever, and test on a mid-range phone over mobile data, where heavy images, slow scripts and pop-ups hurt most. Check field data in Search Console’s Core Web Vitals report or PageSpeed Insights, not only a lab score.

Most small business sites never need to think about it. Google’s guide is aimed at sites with over a million pages that change weekly, sites with over 10,000 pages that change daily, and sites with many URLs stuck in Discovered – currently not indexed.

Where it does matter, the fixes are about wasting less:

• consolidate duplicate URLs and use robots.txt to block endless filter or internal search URLs;
• return 404 or 410 for pages that are gone, and avoid long redirect chains;
• keep sitemaps current and pages quick to load.

Adding noindex does not save budget, because Google still has to fetch the page to see it. For Google Search, the crawling a site gets also reflects its popularity, overall user value, content uniqueness and serving capacity.

Having a few new URLs in this state is normal. Hundreds or thousands often mean Google doesn’t yet think the pages are worth fetching. For Search, Google weighs popularity, overall user value, content uniqueness and serving capacity, and John Mueller said in July 2026 that when its systems are seriously worried about a site’s quality, they reduce the number of pages they index.

What actually helps:

• link to important pages from your menus and from strong existing pages;
• cut or merge thin, near-duplicate pages, such as city or tag pages with nothing unique on them;
• earn real demand: mentions, links and people searching for you by name.

Use “Request indexing” in URL Inspection sparingly, for your most important pages; it does not fix the underlying cause. See crawl budget.

Google’s examples include “multiple domain names or pages targeted at specific regions or cities that funnel users to one page”, and “substantially similar pages that are closer to search results than a clearly defined, browseable hierarchy”.

Location pages are fine when each one is genuinely about that place. The test is whether a visitor from that town learns something they couldn’t get from your other pages:

• a real office, staff or service area there, with directions or contact details;
• local proof, such as projects completed, photos, reviews and prices for that area;
• a person who has checked every page before it goes live.

If you can’t say something true and specific about a town, don’t make a page for it. One honest service-area page is safer than many thin copies. See also programmatic SEO.

There is no markup that opts you in. You can only opt out: nosnippet or data-nosnippet blocks snippets, and a low max-snippet value makes a featured snippet unlikely because there isn’t enough text to show.

What tends to earn one is the same habit that helps with AI answers: a question-style heading followed by a direct, self-contained answer, then a short list or table when the answer is a set of steps or a comparison.

A snippet can mean fewer clicks when the extract fully answers the question. For a business, the best targets are questions where the reader still needs you afterwards, for a quote, a booking or the detail. See zero-click search for that trade-off.

The rule that trips up many sites: each version must list itself and every other version, and the links must be mutual. If two pages don’t both point to each other, Google ignores the tags.

Codes are an ISO 639-1 language, optionally followed by an ISO 3166-1 region: en-IN, ml-IN or ta-IN for India, en-AE for the UAE. A region on its own, such as IN, is invalid. Add x-default for visitors who match none of your versions.

Google works out a page’s language itself; it doesn’t rely on hreflang or the lang attribute for that. A Surat textile exporter with English pages for the UAE and the US can use en-AE and en-US so each market sees its own prices, delivery terms and contacts.

Their weight in local rankings has faded. Whitespark’s 2026 survey of local search experts says citations have been “decreasing in importance for the past decade”, yet the same report rates three of the top five AI search visibility factors as citation factors. That is expert opinion, not measured data. Google itself says only that prominence draws on things like how many websites link to your business and how many reviews you have.

So focus on quality over quantity:

• the main platforms in your market, such as Justdial, IndiaMART or Sulekha in India;
• your industry body, trade associations and chamber of commerce;
• identical details everywhere (see NAP consistency).

Skip packages that submit you to hundreds of low-quality directories.

Prominence is how well known you are, including how many websites link to you and how many reviews you have. You can’t change distance, so the levers are a complete, accurate Business Profile, the right primary category, real photos, a steady flow of reviews with replies, and a website that matches the listing.

The same data also feeds AI answers. In an August 2026 US study of nearly 4,000 local searches per surface, 79.8% of AI Mode’s local citations pointed to Google Maps or Business Profile data, while AI Overviews mostly cited businesses’ own websites. Ask Maps, Gemini’s conversational search inside Google Maps, began rolling out in the US and India in March 2026. Our Business Profile guide covers the details.

Two large infrastructure providers made it easy in February 2026. Cloudflare converts HTML to Markdown at the edge for sites that switch the feature on, adding a header that reports the token count. Vercel published a content-negotiation pattern and showed one page shrinking from about 500 KB of HTML to 3 KB of Markdown.

Keep expectations modest. It is aimed at coding assistants and other agents that read pages directly, but Google says you don’t need Markdown or AI text files to appear in Google Search, and there is little public evidence that it changes citations in other AI tools. If you offer it, keep the Markdown identical in substance to the HTML page and treat it as a companion to llms.txt, not a ranking tactic.

Google’s own rule is the starting point: your profile name should reflect your real-world name “as used consistently on your storefront, website, stationery, and as known to customers”, and adding unnecessary information to it can get the profile suspended. Google doesn’t publish NAP consistency as a ranking factor in itself; the benefit is fewer conflicting records about you.

Practical checks:

• one spelling of the name and one address format, matching your signboard;
• the same main phone number, with +91, on your website and every profile;
• change everything in the same week when you move.

A Kochi clinic moving from MG Road to Kakkanad should update its Business Profile, website footer and directory listings such as Justdial or Practo together, then check them again a month later.

Google lists a passage ranking system among its ranking systems: it identifies individual sections of a page to judge how relevant the page is to a search. AI answers take the idea further, typically retrieving passages for each sub-question in a query fan-out and writing the answer from the strongest ones.

You don’t need to chop pages into fragments. Google says there’s no requirement to break content into tiny pieces, and that its systems can pick out the relevant part of a page that covers several topics. What helps is ordinary clarity: descriptive headings, and a first sentence under each one that answers on its own and names the thing instead of saying “it”. A Kochi clinic’s page on knee replacement written that way gives separate, quotable answers on recovery time and on cost.

For a site owner, the useful part is the deeplink Google documents: https://www.google.com/preferences/source?q= followed by your site’s URL opens the preference screen with your site filled in, so you can put it behind a button in articles, emails or newsletters.

The feature launched in the US and India in August 2025 and now covers Top Stories in every language where Google Search is available. Only whole domains and subdomains are eligible, not sub-folders such as /blog. To show as preferred in AI Mode and AI Overviews, your site must also be included in Search generative AI features in Search Console.

It helps with readers who already value you. It is not a substitute for earning visibility with everyone else.

Google’s spam policies don’t ban templates; they target scale without value. Scaled content abuse covers many pages made “for the primary purpose of manipulating search rankings and not helping users”, and doorway abuse covers near-identical city or region pages that funnel visitors to one place. The August 2026 spam update hit heavily programmatic sites, and Google’s John Mueller has said that when its systems are seriously worried about a site’s quality, they index fewer of its pages.

To stay on the right side:

• build each page on real data a visitor needs: stock, specifications, prices or local reviews;
• publish only combinations with genuine demand and something distinct to say;
• have a person review samples from every template, and remove or noindex thin pages.

A Surat textile exporter’s page per fabric with real weights, minimum order quantities and swatch photos is programmatic SEO done well; a page per city with only the name swapped is not.

Google says more reviews and positive ratings can help your local ranking, and that replying shows customers you value their feedback. In BrightLocal’s 2026 survey of US consumers, 81% expected a reply within a week.

The rules are strict. Google’s policy bans paid reviews and any incentive, such as a discount or freebie, for posting a review or removing a negative one. Businesses must not pressure customers to review while on the premises or ask for specific content. In practice:

• ask every customer the same way, using the review link in your Business Profile;
• check for new reviews every few days, and reply to good and bad ones politely, without sharing personal details;
• flag reviews that look fake, such as a burst of one-star ratings from new accounts or complaints about services you don’t offer, instead of arguing in public.

Our guide to removing fake Google reviews covers reporting step by step.

A common trap: a group for a named bot replaces the * group instead of adding to it. If GPTBot has its own group, nothing under User-agent: * applies to it, so restate every Disallow line in each named group.

Three more things worth knowing:

• A blocked URL can still be indexed from links elsewhere and shown without a snippet. To keep a page out of Google, use noindex on a page Google is allowed to crawl, or a password.
• Google reads only user-agent, allow, disallow and sitemap, ignores other lines such as crawl-delay, and stops reading after 500 KiB.
• Google-Extended is a token, not a crawler. Disallowing it stops use of your content for Gemini training and grounding in Gemini apps, but it does not affect Google Search, including AI Overviews.

Some AI agents that fetch pages for a user may not follow robots.txt at all.

Google introduced the policy in March 2024. Its examples include using generative AI to create many pages without adding value, scraping or synonymising other people’s content, stitching pages together, and pages that make little sense to a reader but contain search keywords. Since May 2026 the same policies also cover attempts to manipulate generative AI responses in Google Search.

Enforcement is active: Glenn Gabe’s case studies of the August 2026 spam update describe programmatic and AI-generated sites losing large amounts of visibility. Using AI is not the violation in itself; volume without value is. To stay on the right side, publish only pages you would happily show a customer, with original information or first-hand experience, edited by a person who knows the subject, at a pace you can actually review. See also doorway pages.

Clicks from AI Overviews and AI Mode are not separated out; they sit inside the normal Performance report under the Web search type. So use this report to see which pages Google draws on, and in which markets, then judge value from enquiries and sales.

You can break impressions down by page, country, date and device. Several links from your site in one AI answer count as a single impression in the chart total, and the newest data is preliminary.

The same rollout added a Search generative AI setting that lets a site opt out of these features. Opting out removes the impressions and traffic they bring, but does not affect regular rankings. For a wider view across engines, see AI visibility tracking.

Google’s news system, announced in 2023, looks at how notable a source is for a topic or location, its original reporting and its reputation, such as a local paper reporting on its own area. For a business site, the practical version is to cover your own field thoroughly and stay in it:

• map the questions customers ask before and after buying, and answer each one well;
• link related pages so the subject hangs together;
• show first-hand experience through cases, photos and named authors.

Google’s self-assessment asks whether you are “producing lots of content on many different topics in hopes that some of it might perform well in search results”; that is the opposite. A Kochi physiotherapy clinic writing about knee, back and sports rehabilitation is building authority; adding unrelated travel tips dilutes it. See E-E-A-T.

Chrome opened an origin trial for it from Chrome 149. There are two ways in: add toolname and tooldescription attributes to an existing HTML form (the declarative API), or register tools in JavaScript with document.modelContext.registerTool() (the imperative API).

Google said at I/O in May 2026 that Gemini in Chrome will “soon” support WebMCP. Until agents use it widely, and while the API can still change, treat it as an experiment:

• start with a form that already works well for people;
• write plain, accurate tool descriptions;
• keep a human confirmation step for anything that books, pays or sends data.

Good HTML still matters more today; see agentic browsing.

A common mistake is a lastmod that changes on every build or every day. Google trusts it only when it matches real, significant updates to the main content, structured data or links; a new copyright year doesn’t count. Google ignores <priority> and <changefreq> entirely.

Other basics:

• list only canonical, indexable URLs that return a 200 status;
• keep each file to 50,000 URLs or 50 MB uncompressed, and use a sitemap index beyond that;
• submit it in Search Console and reference it in robots.txt.

A sitemap invites crawling; it does not guarantee indexing. Bing and several other search engines can also be told about changes straight away through IndexNow.