AEO tooling in 2026: 160 products checked against their own pages, sorted by the layer they work on

By Robin Dally, We Are All Connected. Prices and product claims were read on 6 and 8 September 2026 and will move; the vendor links are there so you can check. We have not trialled the products graded here. The two we run ourselves, Ahrefs Brand Radar and Canonry, are named as such where they appear. The research behind this ran across three AI-assisted passes and Claude wrote the first draft from that evidence; the verdicts, the grades, and every edit are mine, and I am responsible for what is published here.

There is no such thing as an AEO tool. There are seven layers of work, and every product sold as an AEO tool sits on one of them: getting your pages in front of AI crawlers, structuring them, writing them, making sure the engines know which company you are, choosing which prompts to care about, plugging your data into agents, and finding out whether any of it worked. We read the pages of 160 products, what OpenAI, Anthropic, Google, and Microsoft say a site needs, and every study we could find. Two things stood out. The layers with most of the products and most of the marketing, measurement and content scoring, are the two where the vendors' own pages either admit the numbers are modelled or say nothing about method at all. And most of what the platforms say actually matters is free.

If you want the short version: run the free stack first, buy one tracker as an instrument rather than a lever, pay for a content editor as an editor, and pay nobody for schema, llms.txt, or an AI score.

Why there is no AEO tool, only seven layers

Ask a vendor what an AEO tool does and you get one of two answers. It tells you whether ChatGPT mentions you, or it scores your content for "AI readiness". Both are one layer of a seven-layer job.

The way we look at a page has four layers. Can the engines reach it (eligibility)? Can they lift a passage out of it (extractability, which for tools splits into structure and content)? Can they attach the right company to it (attribution and entities)? Is what they lift safe to summarise? Around those four sit three more that tools now sell into: planning, which prompts to write for; the agent layer, how assistants connect to your site and your data; and measurement, whether any of it worked. This guide puts every product we found on one of those layers, prints what the platforms say about that layer, and gives a grade: use, hygiene, unproven, or folklore.

The grades are ours. Nobody has independently tested any of these products, so no grade rests on a test. Three research passes and a red-team search turned up a vendor's own accuracy table and an agency roundup with no control group, and the one academic survey of the field finds no technique that holds up across platforms, never mind a tested product. That is the most important sentence in this guide. We come back to it in the content section.

How we checked 160 products, and what we did not do

Three research passes on 6 and 8 September 2026 read the pages of every product we could name. We read the pricing page or the documentation rather than the homepage, because homepages are marketing rendered by JavaScript and pricing pages are where the numbers are. A second reader re-fetched every page and checked every quote against it. The result is a directory of 160 vendor-product families, meaning one row per vendor and product line with variant names folded in. Of those, 137 have a verified row from the vendor's own page, four were read in a browser only because the fetcher was refused, and 19 are named in our notes and were not read. A further 36 names that came up along the way, crawler user agents, protocols, marketplaces, review sites, and roundup articles, are listed separately and not counted as tools. The full directory is the appendix.

What we did not do is run the tools. Everything below is what the documentation said on the date given. Where a vendor's page and its help centre disagree, we print both. Where a price could only be seen in a browser, it says browser-read. And 160 is a count of what we checked, not of what exists: one roundup we read in the last pass named five optimisation products that appear nowhere in our 160. If you sell one we have not read, the submission route is at the end.

How to read an AEO vendor page in five tests

Look for the admission against interest. A vendor conceding that its number is modelled is telling you something that costs it a sale, so believe it. Ahrefs' methodology page says Brand Radar's headline metrics, AI Share of Voice and Estimated Impressions, are modelled visibility signals rather than measured performance data. Semrush says its topic-level prompt volume is an estimate, built by grouping similar prompts because individual prompts are too specific to count. Otterly built an Estimated Intent Score because OpenAI does not publish prompt volumes. Peec's documentation describes its Prompt Volume score as the output of a model, not a count. These are the pages to trust.

Print "estimate" next to every volume. Unless the vendor says how it counted, it modelled. Profound's feature page says its Prompt Volumes product gives exact keyword volumes. Its FAQ says the data starts with real prompts from opt-in consumer panels and is then corrected by statistical modelling. Its help centre calls it a licensed, modelled dataset. Three pages, three descriptions of the same number. The help centre wins, and the word is "modelled".

Ask which surface it samples. Consumer interface or model API? Profound's comparison page says it runs prompts daily through the front-end browser interfaces of the AI platforms, and that Peec does the same. SE Ranking says it samples the platforms the way a real user would see them. DataForSEO sells both routes as separate products, an API for model behaviour and a scraper for what users see. Neither route gives you a signed-in user. OpenAI's own documentation says ChatGPT search rewrites the user's query into one or more targeted queries before it searches, and passes the user's rough location to its search partners. In August 2026 an agency replayed 62 prompts through four OpenAI API model variants and through the ChatGPT product, and found the API's behaviour did not carry over; their reading is that an API benchmark tells you about a model and its tool contract, not about what someone on chatgpt.com sees. So before you compare two vendors' numbers, ask which surface each one asked, per engine.

Never take a rival's price from a comparison table. On one day in September, three tables gave Peec three different entry prices: 89 euros a month in one, $180 a month in another, $95 a month in a third. Peec's own pricing page, rendered in a browser, showed 70 euros a month on annual billing. Every price in this guide comes from the vendor's own page, with the date and the billing basis.

Ask about the terms of service. OpenAI's Terms of Use say a user may not automatically or programmatically extract data or Output from its services. One scraping-infrastructure vendor's guide gives the reason trackers sample the interface anyway: it returns citation data the API does not. We are not lawyers and this guide draws no conclusion about any vendor. It is a question to ask before you sign, and we flag it wherever it applies.

Layer 1, eligibility: can the engines reach the page?

This is the one layer where the platforms say, in their own words, what they need. OpenAI's developer documentation says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they may still appear as navigational links, and that allowing OAI-SearchBot for search is a separate robots.txt setting from allowing GPTBot for training. Anthropic's help centre says disabling Claude-SearchBot stops its system indexing the site for search, which may reduce the site's visibility in search results. Microsoft's Bing Webmaster blog says Bing respects content-owner preferences for AI citation expressed through robots.txt. Google-Extended governs training and Gemini grounding, has no user agent of its own, and is not a ranking signal. The thing most robots.txt files miss: training bots and retrieval bots are now separate switches, so you can block GPTBot and still allow OAI-SearchBot. Jon's piece on publisher controls covers the mechanics.

Two cautions. The switch leaks. An analysis of the top 50 AI-blocking news publishers found that 82.4% of the sites blocking OAI-SearchBot still turned up in the dataset's AI citations, and that most of the ChatGPT citations in the dataset came from sites that block the retrieval bots. The population is news publishers, so treat the bot setting as a stated policy rather than a proven switch. On Google's side the evidence runs the other way: a SIGIR 2026 study found sites blocking Google's AI crawler were significantly less likely to be retrieved by AI Overviews even with accessible content. And rendering. Vercel's December 2024 analysis with MERJ found that none of the major AI crawlers run JavaScript, with Gemini the exception because it rides on Googlebot's infrastructure. If your important text arrives client-side, the non-Google engines never see it.

The tools on this layer are mostly things you already have.

Cloudflare AI Crawl Control and Vercel's AI bots ruleset

(Free, on all plans)

Cloudflare's AI Crawl Control is on every Cloudflare plan at no extra cost; Vercel's AI bots managed ruleset is on every plan too. Between them they cover most B2B sites, and they tell you which bots fetched what. Cloudflare's Pay Per Crawl, which lets a site charge crawlers per request with HTTP 402 responses, is in closed beta, and AWS WAF has a similar collection mechanism. Those are for publishers with content to sell, not for a B2B site that wants to be cited.

Screaming Frog Log File Analyser

(Free and paid)

The free version handles 1,000 log events and a licence is £99 a year. If you want to know which AI bots are reading your site, this is the cheapest honest answer. The enterprise crawlers do the same job at enterprise prices: Oncrawl separates AI user fetchers from AI search and training crawlers, and Botify and Lumar track named AI bots from customer logs, with no price on the pages we read.

Pixelmojo AI Crawl Checker, MRS Digital's checker, and Siftly's audit tool

(Free)

Three free checkers that tell you whether the named AI bots can reach your pages. Pixelmojo's checks 14 user agents, MRS Digital's shows whether GPTBot, ClaudeBot, and PerplexityBot can get in, and Siftly's audits access for the assistants it names. Run one; the results should match your robots.txt, and if they do not, that is the finding.

Prerender.io

(Paid)

We recommended it in 2017 for search crawlers that could not handle JavaScript. It is back for the same reason with a different set of crawlers, from $49 a month for 25,000 renders. If you cannot ship server-rendered HTML, this is the patch.

• Cloudflare AI Crawl Control. What it does, per its page: Control AI crawler access; Pay Per Crawl in closed beta. Price (read 6 or 8 Sep 2026): No extra cost on all plans.
• Vercel AI bots managed ruleset. What it does, per its page: Control training, search, and user-fetch bots. Price (read 6 or 8 Sep 2026): All plans.
• Oncrawl, Botify, Lumar log tools. What it does, per its page: Which AI bots fetched which pages. Price (read 6 or 8 Sep 2026): Botify and Lumar publish no price on the pages read; Oncrawl's price not read.
• Screaming Frog Log File Analyser. What it does, per its page: Search engine and AI bot crawl analysis. Price (read 6 or 8 Sep 2026): Free to 1,000 events; £99 a year.
• Pixelmojo, MRS Digital, Siftly checkers. What it does, per its page: Whether named AI bots can reach the site. Price (read 6 or 8 Sep 2026): Free.
• Prerender.io. What it does, per its page: Server-side rendering for AI crawlers. Price (read 6 or 8 Sep 2026): From $49 a month, 25,000 renders.
• ai.robots.txt. What it does, per its page: Community blocklist of AI crawlers. Price (read 6 or 8 Sep 2026): Free, MIT licence.

One piece of folklore lives on this layer: llms.txt as a lever. The agent layer section takes it properly. The short version is that Google's guidance tells sites to ignore AEO hacks such as creating AI text files, and a 137,000-domain log study found the files almost never fetched.

Layer 2, structure: does schema help AI search?

Google's AI features documentation says there are no additional requirements and no special structured data for appearing in AI Overviews or AI Mode; SEO fundamentals apply, and the nosnippet, data-nosnippet, max-snippet, and noindex controls decide what AI features can show. Its structured-data documentation says markup makes a page eligible for a defined list of rich-result features, and not for AI answers in general. That is the whole of the platform position, and it is Google's only; no other engine has published anything like it.

The two tests we found agree outside Google and disagree inside it. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched control pages. It found no meaningful uplift in AI citations on any of the three platforms tested, and on Google AI Overviews the schema-treated pages fell 4.6% relative to controls, a small but statistically significant decline. Otterly put five schema types on its own site and tracked 319 prompts across seven platforms for three months. Most platforms could not read the markup when asked, only Gemini returned correct JSON-LD, Google AI Mode hallucinated schema types the page did not have, and the gains Otterly reported were inside Google's own products. One matched-control study says down inside Google; one single-site study says up. We print both and trust the matched design more. The one academic paper on the question sits behind a challenge page we could not read.

So ship schema that matches the page and validate it, because it makes you eligible for rich results and costs almost nothing. Pay nobody for schema as an AI lever.

Rich Results Test and the Schema Markup Validator

(Free)

Same advice as our SEO tools review, which lists both as free: the Rich Results Test for eligibility, the Schema.org validator for whether the markup is well formed. AdNabu's checker calls itself 100% free with no sign-up, and Nuxt SEO's validator checks one page at a time for nothing. On WordPress, Yoast outputs a schema graph by default and Rank Math supports schema natively, so you already have this layer.

• Rich Results Test, Schema.org validator. What it does, per its page: Validate markup. Price: Free.
• AdNabu Schema Checker, Nuxt SEO validator. What it does, per its page: Validate one page at a time. Price: Free.
• Yoast SEO, Rank Math. What it does, per its page: Schema graph built in. Price: Included.
• Schema App, WordLift, InLinks. What it does, per its page: Entity-linked schema platforms. Price: Schema App by contact; WordLift Business+ 799 euros a month yearly or 999 monthly; InLinks from $49 a month.

Layer 3, content optimisers: do AI content scores work?

Start with what is known, because the idea is not the problem. Content does shape citation. A KDD 2024 paper showed visibility gains of up to 40% in generative-engine responses from defined content changes. A May 2026 factorial trial ran 252,000 trials across six models on a two-document testbed and found that topical relevance and list position are the biggest causal drivers of being cited first, that explicit prices and a recent timestamp help, and that formatting-only edits do very little, contrary to the "structure your content" advice you will have read everywhere. A structural feature-engineering framework reported a 17.3% citation-rate improvement. Then read the July 2026 critical survey of 45 GEO studies. Content that has already been retrieved can be shown to change its own citation, but no reviewed technique shows a stable, long-run, cross-platform effect on being found in the first place; relevance and position are the most reproducible levers; heuristics transfer badly between platforms; and citation-oriented rewrites can make retrieval worse. The mechanism is real, narrow, and platform-specific.

Now the products. Frase scores drafts on SEO and GEO and bundles an AI Visibility feature. Surfer sells a standalone AI Tracker with a Visibility Score and share of voice, and says nothing about method on the pricing page. Clearscope's Prompt Tracking follows prompts across ChatGPT and Gemini. NeuronWriter ties its scoring to AI search credibility. Every optimiser we read now leads with prompt tracking, which is measurement wearing a content label. None publishes a method for its AI score. The before-and-after evidence is the vendors' own case studies. Profound's Ramp story reports AI visibility rising from 3.2% to 22.2% in a month, puts it down to two new pages that drew over 300 citations, and gives no prompt set, engines, or run counts. Otterly's medical-device story headlines an 8x increase over 12 months while its body reports 2,016% and 2,670% year on year, and nowhere reconciles the two. AthenaHQ attributes a Gruns share-of-voice lift from 2.0% to 12.6% in 60 days to content written through Athena. Read the Ramp story closely. The tool was the diagnostic; the new pages were the intervention. That is content plus measurement, sold as a tool effect.

We searched three times, and a red team searched a fourth time, for a test of any named optimiser with an untreated control, a stated sample size, and a spread. There is none. What exists is a vendor's accuracy table that ranks the vendor first among the competitors it scores and an agency roundup that publishes per-tool citation gains with no control group. So "unproven" is the honest grade, and it applies to the products, not to the idea. A tool that works and does not publish its method gets the same grade, which is a limit of reading documentation rather than a finding against the tool.

What to buy is an editor you like, at an editor's price.

Frase

(Paid)

Starter is $39 a month billed yearly or $49 monthly. The GEO scoring and the AI Visibility feature come bundled; treat them as free extras.

Clearscope

(Paid)

Essentials is $129 a month, with Prompt Tracking across ChatGPT and Gemini included. Of the three, the AI claim on the page is the smallest.

Surfer

(Paid)

The standalone AI Search Analytics product is $158 a month billed yearly. Its AI Tracker sells a Visibility Score with no method on the page, which is the thing to price at zero.

• Frase. What its page claims: SEO and GEO scoring, AI Visibility feature. Price (read 6 or 8 Sep 2026): $39 a month yearly, $49 monthly.
• Surfer. What its page claims: AI Tracker, Visibility Score, no method on page. Price (read 6 or 8 Sep 2026): $158 a month, billed yearly.
• Clearscope. What its page claims: Prompt Tracking across ChatGPT and Gemini. Price (read 6 or 8 Sep 2026): Essentials $129 a month.
• NeuronWriter. What its page claims: Content scoring for AI search credibility. Price (read 6 or 8 Sep 2026): See page.
• AthenaHQ. What its page claims: Prompt-led content, share-of-voice case studies. Price (read 6 or 8 Sep 2026): Starter $295 a month, 3,600 credits, 10 models.
• Scalenut. What its page claims: AI visibility system plus GEO content. Price (read 6 or 8 Sep 2026): Starter $59 a month list, $24 a month promotional.
• Keytomic. What its page claims: Automated research, publishing, tracking. Price (read 6 or 8 Sep 2026): Pro $99 a month (homepage shows it as discounted from $299; pricing page shows $99 flat and an Agency plan at $999).
• Gauge. What its page claims: 600 prompts daily plus 18 articles a month. Price (read 6 or 8 Sep 2026): Growth $599 a month.
• AEO Engine. What its page claims: AEO content service. Price (read 6 or 8 Sep 2026): Growth $1,597 a month, month to month.

Layer 4, entities: can you buy your way into the knowledge graph?

The knowledge graph is where an engine turns your company name into a thing it knows about. Google's Knowledge Graph Search API gives 100,000 free read calls a day, and Google's own page says it is read-only: a way to see what Google holds about you, not a way to change it. Wikidata accepts an item only if it has a sitelink, is a clearly identifiable entity described by serious public references, or fills a structural need. Its guidance strongly discourages an organisation creating an item about itself and treats paid promotion as self-promotion with disclosure rules, and it says personal websites, blogs, press releases, and marketing materials do not count as references, which is the reference base most mid-market B2B companies have. Wikidata is a gate, not a tool.

What you can act on is the surfaces the engines cite. In Similarweb's study of nearly 600,000 ChatGPT citation events, Wikipedia and Reddit each took roughly 12 to 13% of web-browsing-mode citations, and LinkedIn sat in the top 20 domains at 2.42%. A basic G2 vendor profile is free indefinitely. The one number linking profile work to citations is G2's own study on Profound's data: 500 categories, 30,000 citations, a small but statistically reliable positive relationship between review volume and citation share. It is published by the marketplace that sells the reviews, on a tracker vendor's data, and it explains under 1% of the variance. It is the only number there is, and on its own it is not a reason.

The paid entity platforms publish no effect evidence and mostly no price: Yext's Knowledge Graph is sales-gated, Kalicube funnels you to a free audit, and Schema App's entity linking is by contact. Treat them as consultancies with software attached.

• Google Knowledge Graph Search API. What it is: Read-only check of Google's entity data. Price: Free, 100,000 calls a day.
• Wikidata. What it is: A gate with notability rules, not a tool. Price: Free; self-promotion discouraged.
• G2 vendor profile. What it is: Cited surface; basic profile free. Price: Free.
• LinkedIn company page. What it is: Cited surface. Price: Free.
• Yext Knowledge Graph, Kalicube, Schema App. What it is: Paid entity management. Price: No price on page.

Layer 5, planning: are AI prompt volumes real numbers?

Prompt research is where the incumbents' keyword corpora earn their money, and it is also where the most confident numbers are the least counted. Semrush's knowledge base defines AI Volume as an estimate of the queries a topic gets on a platform, and its Brand Performance reports run on Semrush-generated synthetic prompts, a mix of branded and non-branded queries rather than anything a real user typed. Ahrefs builds Brand Radar's prompt indexes by taking real queries from its keyword database, expanding them through People Also Ask and semantic fan-out, and running the results through each platform; its default AI Chatbots index refreshes once a month, and custom prompts can be checked daily to monthly. Profound's Prompt Volumes is a licensed, modelled dataset by its help centre's own description. Only Similarweb claims to show the actual questions users type. Treat that outlier with the same scepticism as Profound's "exact".

Volatility is the reason to use these tools and the reason not to trust their precision. A January 2026 study of 2,961 AI responses found under a one-in-a-hundred chance that ChatGPT or Google's AI would return the same list of brands twice for the same prompt. A prompt set is a sample of a moving target. A modelled volume tells you which topics are worth sampling, and nothing more exact than that.

Otterly prompt research

(Included in Otterly plans)

Turns a seed topic into hundreds of prompts grouped by intent and funnel stage, with a query fan-out view.

Gumshoe and SparkToro

(Paid)

Gumshoe simulates AI conversations with personas rather than scraping raw outputs. SparkToro includes "AI prompts" as an audience data type in every paid plan. Both are for finding the language, not for measuring it.

AlsoAsked

(Free and paid)

Basic is $12 a month or $144 a year for 100 credits, and it is still the quickest way to see the questions around a topic before you write.

One caution: edit any generated prompt set before you track it, and read each prompt as a buyer would say it.

• Ahrefs Brand Radar prompt index. What it does, per its page: Keyword-derived prompts run through each platform; monthly index refresh. Price (read 6 or 8 Sep 2026): See measurement table.
• Semrush Prompt Research. What it does, per its page: Keyword research for the AI era; AI Volume is an estimate. Price (read 6 or 8 Sep 2026): In the AI Visibility Toolkit.
• Otterly prompt research. What it does, per its page: Seed topic to hundreds of prompts, with fan-out view. Price (read 6 or 8 Sep 2026): In Otterly plans.
• Gumshoe. What it does, per its page: Persona-simulated conversations. Price (read 6 or 8 Sep 2026): See page.
• SparkToro. What it does, per its page: Audience research with AI prompts data type. Price (read 6 or 8 Sep 2026): All paid plans.
• AlsoAsked. What it does, per its page: Question mining. Price (read 6 or 8 Sep 2026): $12 a month or $144 a year, 100 credits.
• Similarweb Prompt Analysis. What it does, per its page: Claims actual user questions. Price (read 6 or 8 Sep 2026): Contact sales.

Layer 6, the agent layer: connectors, storefronts, and the llms.txt question

The fastest-growing thing on this layer is not a product but a plug. MCP servers let an assistant read a tool's data directly, and most of the tools you already pay for now have one. Ahrefs' remote MCP server is on paid plans from Lite; Semrush's needs a plan with the Standard API allowance; Similarweb's needs API access on its API-only, Business, or Enterprise plans; DataForSEO's runs through its normal pay-per-unit account with no separate fee; Otterly's connects Claude, ChatGPT, and Cursor to its visibility data; Nightwatch includes its MCP server on every plan including Starter. The test for any new dashboard product is whether it does something a connector plus the assistant you already use does not.

On the site side, the job is readiness. Chrome Lighthouse 13.3.0 added an agentic browsing category to its default audit, and that is the free check to run. If you sell through a cart, Shopify's Agentic Storefronts puts your catalogue into ChatGPT, Gemini, and other AI chats, structured automatically through Shopify Catalog, switched on for eligible stores subject to terms, with no separate price; the help centre adds Google AI Mode, Copilot, and Meta to the list. OpenAI's Agentic Commerce Protocol is open to build against, while Instant Checkout inside ChatGPT is restricted to approved partners and needs three flows from a merchant. For a B2B site with no cart, none of this is a purchase decision yet.

Then llms.txt, which we have to deal with because every CMS now makes one. Yoast generates it in one click for free, Rank Math too, Wix for upgraded custom-domain sites, the docs platforms Mintlify, Fern, GitBook, and ReadMe by default, and WordPress plugins such as LLMagnet. The evidence that anyone reads the file: Ahrefs' June 2026 log study across 137,000 domains found that publishing one almost always results in no fetch at all. Redocly's CEO wrote that after building automatic support and testing across models, no model read or respected the file on its own, and that server logs show it is basically never accessed. A generator vendor's own FAQ says the 2026 evidence shows no measurable effect on AI search rankings. And Ahrefs reports that Google's generative-AI guidance says, under a mythbusting heading, that machine-readable files such as llms.txt are not needed to appear in generative AI search. Publishing one costs nothing and we publish one ourselves as cheap insurance. Paying for one, or putting it in a report as AEO work, is folklore.

"Agentic search optimisation" as a product category has no measured outcome anywhere we could find. Adobe names it as a third layer beside SEO and GEO, and a competitor's blog calls most of what is sold under the label data hygiene with a new name. No evidence either way, so unproven.

The CMS and plugin layer is real but small. Webflow AEO measures answer-engine presence and ships changes inside Webflow, available with Analyze for Enterprise and priced by sales. On WordPress: Geoa builds registry-verified schema and llms.txt with before-and-after proof; Visibility, by Fernando Tellado, covers SEO essentials with no Pro tier and no upsells; Opttab's free plugin gives a sitewide GEO score, with paid plans on its own site from $84 a month on annual billing. Check the active-install count on the listing before you put any of them on a live site.

• Ahrefs, Semrush, Similarweb, DataForSEO, Otterly, Nightwatch MCP servers. What it does, per its page: Assistant access to the tool's data. Price (read 6 or 8 Sep 2026): Included in plans as stated.
• Chrome Lighthouse 13.3. What it does, per its page: Agentic browsing audit. Price (read 6 or 8 Sep 2026): Free.
• Shopify Agentic Storefronts. What it does, per its page: Catalogue into AI chats; on by default for eligible stores. Price (read 6 or 8 Sep 2026): No separate price.
• OpenAI Agentic Commerce Protocol, Instant Checkout. What it does, per its page: Open to build; checkout restricted to partners. Price (read 6 or 8 Sep 2026): Partner-gated.
• Yoast, Rank Math, Wix, Mintlify, Fern, GitBook, ReadMe. What it does, per its page: llms.txt generation. Price (read 6 or 8 Sep 2026): Free or included.
• Webflow AEO. What it does, per its page: Measure and ship changes inside Webflow. Price (read 6 or 8 Sep 2026): Enterprise, by sales.
• Geoa, Visibility, Opttab, LLMagnet. What it does, per its page: WordPress AEO plugins. Price (read 6 or 8 Sep 2026): Free tiers; Opttab from $84 a month annual.

AI search advertising: who is paying to appear next to you

This is a category to watch rather than a layer with a grade. OpenAI's ChatGPT ads target on the context of the current conversation, and self-serve access opened to Europe, India, the Middle East, and North Africa on 31 August 2026, with no per-unit price on the page. Perplexity started experimenting with sponsored follow-up questions in November 2024. What matters for an organic programme is that the trackers you buy for citations now show you the paid placements beside them. Otterly's ChatGPT Ads Tracking records the advertiser, destination, copy, and product behind paid placements and Shopping Cards, included in every Brand Report at no extra cost. Am I Cited sells a ChatGPT Ads Manager that measures campaigns against organic citations. SE Ranking's tracker page reports its own research finding paid placements on roughly one in four commercial ChatGPT prompts, and argues that mentions alone do not tell you who is paying. Adthena's ChatGPT Ads Intelligence is $399 a month with a 21-day trial and sells itself as showing what ChatGPT's own advertiser reporting leaves out.

• OpenAI ChatGPT ads. What it does, per its page: Conversation-context targeting; self-serve in Europe from 31 Aug 2026. Price (read 8 Sep 2026): No per-unit price shown.
• Otterly ChatGPT Ads Tracking. What it does, per its page: Advertiser, destination, copy behind placements. Price (read 8 Sep 2026): Included in Brand Reports.
• Am I Cited Ads Manager. What it does, per its page: Campaigns measured against organic citations. Price (read 8 Sep 2026): No price on page.
• SE Ranking ChatGPT Ads Tracker. What it does, per its page: Paid placement tracking. Price (read 8 Sep 2026): No standalone price on page.
• Adthena ChatGPT Ads Intelligence. What it does, per its page: Competitor ad activity in ChatGPT. Price (read 8 Sep 2026): $399 a month; 21-day trial.

Layer 7, measurement: how do you know if any of it worked?

Route one: the free first-party reports, and what they actually count

Google Search Console

(Free)

The generative AI performance report counts impressions of links to your site shown in AI Overviews and AI Mode, combined, with pages, countries, dates, and devices as dimensions, no query dimension, and no clicks. Search Engine Land reported on 31 August 2026 that it is available to every Search Console account. It measures link exposure for pages you own. It cannot tell you whether your brand was mentioned. It is still the best free view of Google AI exposure there is.

Bing Webmaster Tools

(Free)

The AI Performance report, in public preview since February 2026, shows when a site is cited across Copilot and Bing's AI answers, with Average Cited Pages reported per day as a sample. Ten minutes to verify, as it was in 2017.

GA4 and Plausible

(Free and paid)

GA4 has a native AI Assistant channel for arrivals from assistants, and Google's own documentation gives a worked regex recipe for a custom channel group on AI assistant referrers if your property predates it. Plausible groups visits from ChatGPT, Claude, Perplexity, Gemini, and Copilot into one channel and shows pages, sign-ups, and purchases per assistant. Set these up before you buy anything, because they answer the question a buyer asks first: did anyone arrive?

Route two: prompt trackers, and what they tell you about their own method

A tracker asks the engines a fixed set of prompts and records who was mentioned and cited. The number to ask for is how many times it asks. Evertune is the only vendor whose documentation answers. Its methodology page says repeated sampling narrows the margin of error, that 100 samples on one prompt give about a ten-point margin, and that sampling 100 prompts 100 times each tightens it to about one point overall. Its FAQ says it samples every prompt 100 times per model across 11 or more models, and claims, as its own comparison, that most platforms sample each prompt once per model per day. The others tell you how often, not how deep. Otterly monitors daily as a neutral, non-personalised user, admits answers vary run to run, and states no run count or interval. Peec runs prompts daily and looks at patterns over time because answers vary day to day. Semrush updates tracked prompts daily and Brand Performance weekly. Ahrefs refreshes its index monthly and custom prompts daily to monthly. Rankscale sells intervals from bi-hourly to bi-monthly and states no sampling count. Profound's help centre and FAQ pages state no run frequency, sampling depth, or interval for its tracking product anywhere we read. No tracker has been tested for accuracy by anyone independent; the only named-product accuracy table is a vendor's own, with itself first. And the surface question from the tests above applies here in full: the trackers sample the ChatGPT interface, one scraping vendor's guide says that is because the interface returns citation data the API does not, and OpenAI's terms restrict programmatic extraction. Ask.

What they cost, from the vendor's own page. Otterly Lite is $29 a month for 15 prompts across four engines; Radarkit Lite is $29 a month with a 7-day trial; Waikay's Small Teams tier is $69.95 a month for 120 prompts; Menra's entry plan is $69 a month or $662.40 a year; Visby's Starter is $79 a month; Omnia's Growth tier is 79 euros a month; PromptWatch's Essential is $95 a month for 50 prompts and four models; Semrush's AI Visibility Toolkit is $99 a month per domain billed annually for 25 prompts; Searchable's Pro is $125 a month for 100 prompts across three engines; AIclicks runs $59, $189, and $499 a month; Ahrefs Brand Radar's AI Visibility Index is $199 a month for a single platform index, or $699 a month for all platforms; its product page attaches 2,500 custom checks a month to the $199 tier and its help centre attaches them to the $699 tier, so ask which before you buy; Scrunch's Core plan is $250 a month. Peec prices by tracked prompts, 50, 150, or 350 per plan across three models with daily tracking, and its rendered page shows 70 euros a month on annual billing for the entry tier. HubSpot AEO is $50 a month or included in Marketing Hub Pro and Enterprise, and HubSpot's free AI Search Grader is a one-time check with no account.

The enterprise tier prices by conversation. BrightEdge's pricing page is fully sales-gated; Conductor's is usage-based with no dollar figures, metered in AI Search Credits and tracked keywords; seoClarity's three ArcAI packages are all "Ask for a Quote"; Lumar's is tailored per customer; Semrush Enterprise AIO is custom; Botify's AI Visibility page has no price; Similarweb's AI Search Intelligence line has no self-serve price. Yext Scout tracks win rates against a brand's 20 closest competitors, is sales-gated on its product page, and is described by Yext's own help centre as in beta and available only to beta participants. Adobe LLM Optimizer has become Adobe Brand Visibility, combining Adobe's optimisation with Semrush's market intelligence; its documented trial gives eligible Adobe customers 100 prompts, one domain, and 50 URLs, and no list price appears on any Adobe page we could read.

If we could only pick one self-serve tracker on this evidence, we would pick on disclosed sampling depth first and price second. If the budget is $29 a month, buy the cheapest one, treat the numbers as a sample, and never report a single day's reading to anyone.

Route three: CDN logs

ScalePost and TollBit

(Price not published on the pages read; usage-based at TollBit)

ScalePost measures AI citations from first-party CDN data rather than by running prompts, integrating with Fastly, Cloudflare, and Akamai to identify the bot, the URL, and the AI surface, and sells the result to publishers as proof of AI influence and to brands as a map of which publishers drive their visibility. TollBit sits on the same data for licensing negotiations and reports over 22 billion AI bot scrapes detected across its network in the first half of 2026. One caution: a CDN fetch is evidence that an engine retrieved a page, not that it cited it in an answer, and the vendor's use of the word "citation" is a claim to test. For a publisher this looks like the most direct measurement on the market. For a B2B site it is a way to see who fetches you.

• Evertune. Engines and cadence, per its documentation: 11-plus models; 100 samples per prompt per model. Sampling depth disclosed: Yes, with a margin of error. Price (read 6 or 8 Sep 2026): Not on page read.
• Otterly. Engines and cadence, per its documentation: Daily, four engines on Lite, add-ons for more. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): Lite $29 a month, 15 prompts.
• Peec. Engines and cadence, per its documentation: Daily, three models per plan. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): 50, 150, or 350 prompts per plan; entry 70 euros a month annual.
• Semrush AI Visibility Toolkit. Engines and cadence, per its documentation: Prompts daily, Brand Performance weekly. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): $99 a month per domain, annual, 25 prompts.
• Ahrefs Brand Radar. Engines and cadence, per its documentation: Index monthly; custom prompts daily to monthly. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): $199 a month single index; $699 all platforms.
• Rankscale. Engines and cadence, per its documentation: Bi-hourly to bi-monthly intervals. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): See page.
• Profound. Engines and cadence, per its documentation: Daily via the interface, per its comparison page. Sampling depth disclosed: Not stated on any page read. Price (read 6 or 8 Sep 2026): See vendor.
• Scrunch. Engines and cadence, per its documentation: See page. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): Core $250 a month.
• HubSpot AEO. Engines and cadence, per its documentation: See page. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): $50 a month or included in Marketing Hub Pro.
• Searchable, PromptWatch, Waikay, Menra, Visby, Omnia, Radarkit, AIclicks. Engines and cadence, per its documentation: Daily trackers on entry tiers. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): $29 to $125 a month as listed above.
• BrightEdge, Conductor, seoClarity, Lumar, Botify, Semrush Enterprise, Similarweb, Yext Scout, Adobe Brand Visibility. Engines and cadence, per its documentation: Enterprise AI visibility modules. Sampling depth disclosed: No. Price (read 6 or 8 Sep 2026): By conversation; no published price.
• ScalePost, TollBit. Engines and cadence, per its documentation: CDN-log measurement of AI fetches. Sampling depth disclosed: Not applicable. Price (read 6 or 8 Sep 2026): ScalePost price not read; TollBit usage-based, no fixed price.

Should you build your own tracker?

The parts are cheap and public. Anthropic's web search tool on the Claude API costs $10 per 1,000 searches plus token costs. Google's Gemini grounding costs $14 per 1,000 queries beyond 5,000 free a month on Gemini 3.x, and the Developer API page confirms the same figures and adds that one request may run more than one search. DataForSEO sells LLM response data and citation metrics across ChatGPT, Gemini, Google AI Overview, Claude, and Perplexity on pay-as-you-go pricing; we measured one browsing-enabled ChatGPT call at $0.027. Canonry is an open-source, self-hosted platform that tracks what ChatGPT, Claude, Gemini, and Perplexity say about a brand, keeps the answers locally, and exposes MCP tools, under a licence that turns into Apache 2.0 after two years.

The catch is the surface question again. The API route measures a model under API conditions and a tool contract. The product a buyer types into rewrites the query, knows roughly where they are, and does not behave like the API on the same prompts. A home-built tracker on the API is honest, auditable, and cheap, and it measures a different thing from the one your client is asking about. Our own rule, set in May 2026 when we chose Canonry for exactly that auditability, is to keep the build only if the measured spend at the cadence you need comes in under the cheapest comparable vendor tier, and otherwise to buy the feed and keep the build for the method and the raw-answer archive. At our volume the modelled cost of a route-substituted set came to about $75.78 a month, a scenario figure and not a measurement, against self-serve tiers from $29 to $199 a month. The archive is what justifies the build. The saving does not.

The free stack a B2B team can run today

Every item here is first-party, free, or included in something you already pay for, and each one says what it counts.

1. Search Console's generative AI performance report: impressions of links to your pages in AI Overviews and AI Mode, combined, no queries, no clicks.
2. Bing Webmaster Tools' AI Performance report: citations across Copilot and Bing's AI answers, a sample, per day.
3. GA4's AI Assistant channel, or Google's regex recipe for a custom channel: arrivals from assistants. Plausible does the same natively.
4. Cloudflare AI Crawl Control or Vercel's bot ruleset: which AI bots fetch what, at no extra cost.
5. Chrome Lighthouse 13.3's agentic browsing audit on your five most important pages.
6. Google's Rich Results Test and the Schema.org validator on the same five pages.
7. One of the free AI crawler checkers, to confirm GPTBot, OAI-SearchBot, and ClaudeBot can reach you.
8. Adobe's free Chrome extension, which checks what LLMs can and cannot read on a page and funnels to the paid product.
9. HubSpot's free AI Search Grader, a one-time check of what three engines say about your brand.

Between them these cover eligibility, referral attribution, and a first look at brand presence. What they cannot do is answer the brand-mention question over time for a fixed set of prompts. That is the one thing a tracker sells, and it is worth buying as an instrument, priced against the disclosure table above, once the free stack is running.

What We Are All Connected uses, and why

Our own stack is Screaming Frog, Dragon Metrics, Ahrefs, Google Analytics, ClickUp, and Google Drive, with Claude, Gemini, ChatGPT, and NotebookLM as the AI platforms we work in. On our site we ship Organization and WebSite schema on every page and Service, FAQPage, and BreadcrumbList on service pages, and we publish an llms.txt file as cheap insurance while our own research calls it theatre. Our AEO guide names schema-as-AEO, llms.txt, FAQ markup, and magic word counts as folklore to stop paying for, and three research passes since have not moved that.

For measurement we track five daily UK ChatGPT prompts in Ahrefs Brand Radar on the Lite plan, which uses the full 150 checks a month the plan includes, and we run Canonry, self-hosted, because its answers are per question and archived rather than modelled into a score. We publish the method and keep the client models; that is our openness rule, and it is why this guide prints the tests rather than a ranking.

What this guide does not cover, and how to get a tool added

• Nineteen products are named in our notes and were not read (the count is from the directory in the appendix), and one roundup named five more we had never seen. There are more than 160 tools; 160 is how many we checked.
• Two analyst pieces stay unread because they are paywalled: Forrester's AEO Technologies Landscape for Q3 2026 and Kevin Indig's Growth Memo state-of-AEO piece.
• Several pages could not be read by our tools and were not read by hand either: the SSRN paper on schema and AI citations, Crunchbase, Kalicube Pro's page, Profound's run-frequency documentation if it exists, a Bing help-centre page on AI answers, the Apify actor listing for llms.txt generators.
• We have not run any of these tools. Every sentence above is what a page said on 6 or 8 September 2026.
• No independent test of any named product's effect exists, and no independent accuracy test of any tracker exists; the closest things we found are a vendor's own accuracy table and an agency roundup without controls. If you have one with an untreated control, a sample size, and a spread, send it and we will read it and update this page.
• The trackers sample consumer interfaces that the platforms' terms restrict programmatic extraction from. We flag it and do not conclude on it.

To add a tool: send the pricing page or documentation URL, the date, and what the page says about method. We read pages, not pitches.

Common questions

Is an AEO tool different from an SEO tool? Mostly no. Google says there are no additional requirements for AI Overviews or AI Mode beyond standard search fundamentals. The genuinely new products are prompt trackers, AI referral analytics, and the connectors on the agent layer.

How much do AI visibility tools cost? Self-serve trackers run from $29 a month (Otterly Lite, Radarkit) to $199 a month for a single Ahrefs index and $699 for all its platforms; the enterprise platforms publish no price at all.

Ahrefs Brand Radar or Profound? They measure different surfaces on different cadences. Ahrefs refreshes its index monthly from keyword-derived prompts and says its metrics are modelled. Profound samples the ChatGPT interface daily and states no sampling depth on the pages we read. Neither has been tested independently.

Does schema markup help AI search? Not on the evidence. Google says no special markup is needed; Ahrefs' matched test found no uplift on any platform and a small decline in Google AI Overview citations. Ship it for rich results and validate it. Do not pay for it as an AI lever.

Is llms.txt worth adding? It costs nothing, so add it if your CMS makes one. Expect nothing from it: a 137,000-domain log study found the file almost never fetched, and Ahrefs reports Google's guidance saying it is not needed.

Do I need to allow GPTBot, OAI-SearchBot, and ClaudeBot? Allow the search bots if you want to appear in ChatGPT and Claude search answers; OpenAI says opting out of OAI-SearchBot removes you from search answers. Training bots are a separate switch. Blocking is a stated policy, not a proven switch.

How often do trackers check? Daily is the norm, Ahrefs' default index is monthly, and only Evertune says how many times it asks each prompt.

Can I see ChatGPT traffic in GA4? Yes. GA4 has a native AI Assistant channel and Google documents a regex recipe for a custom one.

Does Search Console show AI Overview performance? It shows impressions of links to your pages in AI Overviews and AI Mode together, with no query dimension and no clicks.

Should we build our own tracker? Only for the archive and the method. The API measures a different thing from the product your buyers use, and self-serve tiers start at $29 a month.