Understanding your AI visibility metrics

AVI — your AI Visibility Index

AVI is a 0–100 score summarising how visible your brand is across the engines and prompts you track, weighted by how prominently you're mentioned. Watch the trend, not the absolute number: a rising AVI means AI assistants are recommending you more often.

Named vs sourced

There are two ways to appear in an AI answer: being named as a recommendation, and being cited as a source the answer was built from. Both matter — named drives brand consideration, sourced drives clicks. The Named vs Sourced view shows where you stand on each.

The perception map

The Brand Perception Map places you and your competitors on two axes built from how AI answers talk about each brand. It needs mention data on more than one brand to render, so it fills in after a few runs with competitors configured.

Perception by attribute

Every answer that names a brand is also read for what it says the brand is known for: up to five attributes per brand (price, quality, customer service, speed and so on), each said as a strength, as a weakness or neutrally, and whether the answer raises a concern about the brand. Insights shows the attributes the answers of the last 30 days tie to you and your competitors most often: in how many answers, and in what tone, plus how many answers raise a concern about each brand. The answer page and buyer journey turns show them answer by answer; hover an attribute to see the answer's own words. They come from the same pass that finds position and sentiment, so they cost nothing extra. Answers checked before 28 September 2026 don't have them.

Gaps and Fixes lists what the answers praise you for (said as a strength in two answers or more) that the tracked prompts asking who's best for it don't back up: they name you in fewer than half of their answers, or no tracked prompt asks for it yet. A prompt asks for it when it has at least half of the words the answers use for it. The answers also suggest how a buyer would ask (for example “Which plumber in Ljubljana is best for emergency repairs?”), and managers can track those questions as prompts in one click, free, within the project's prompt limit.

Citations and citation gaps

Asheard tracks which of your URLs engines actually cite — and, more importantly, the citation gaps: source pages where several of your competitors are mentioned and you are not. Those pages are your outreach and content hit-list, listed on Gaps and Fixes.

What the AI searched for and read

Open any answer to see the searches the engine ran and every page it read. Pages it read but didn't cite are the near misses: the engine found them and picked something else. A page the engine never reads can't be cited at all, so if your site is missing there, the fix is getting found for those searches first.

ChatGPT, Gemini, Claude, Perplexity and Grok report both. Google AI Overviews, AI Mode and Bing Copilot show the pages behind their answer but not their searches. DeepSeek and Mistral answer from memory, without searching. Answers checked before 27 September 2026 don't include this detail.

Search queries

The Search queries page (AI visibility → "What the engines searched for") collects every search the engines ran for each prompt over the last 30 days. "Words engines add" shows what they put into your prompts, such as best, reviews, prices or the year: that is what they think your buyers need. A search marked "Your site never came up" is one where none of the answers read your site, so a page matching that search is the gap. A search marked "English" ran in English for a non-English prompt (a word-based guess), which usually means the engine found too little in your language. Track any search as a prompt in one click; it is checked from the next run.

Google and Bing AI reports

Google and Bing now report on AI search themselves. Search Console's generative AI performance report counts impressions of your pages in AI Overviews and AI Mode (without clicks). Bing Webmaster Tools' AI Performance counts how often Bing and Copilot answers cite your site, and lists the searches behind them (grounding queries).

Neither has an API yet, so export them and upload the files on the Google & Bing page (AI visibility → Google & Bing): the Search Console zip as it is, and Bing's overview and grounding queries. Files are recognised by their columns. Bing's grounding queries file has no dates, so enter the period you exported; importing the same days again replaces them. Google lists only its top 1,000 pages and Bing only part of its searches, so the page shows how much of the total they cover.

Your share of a Bing search is your part of all the citations Bing showed for it. Searches where other sites get most citations are listed on Gaps and Fixes, and any of them can be tracked as a prompt. Pages Google shows are matched with the pages the engines we track cited in the last 30 days.

Win rate and the visibility funnel

Win rate is the share of answers naming you that put you first. Visibility tells you whether you're in the answer; win rate tells you whether you're the recommendation.

The visibility funnel on Insights follows each engine's answers through four steps: the engine reads your site, cites it, names you, and puts you first. Where your brand drops out tells you what to fix, and Gaps and Fixes counts the answers lost at each step. Never read: get found (indexing, and mentions on the sites the engine does read). Read but not cited: the page lost, so lead with a direct answer. Cited but not named: say your brand name clearly on the page being quoted.

Cited when read

Of the answers in which an engine read your site, the share that cited it: the funnel's step from read to cited, as a rate. Insights shows it per engine next to the brands tracked in other Asheard workspaces: the middle half of their rates over the last 30 days (hover for the median). A rate below most brands means the engine finds your pages and then quotes other sources, so lead with a direct, quotable answer; Gaps and Fixes lists those engines.

To keep the comparison fair, both sides leave out branded and comparison prompts (an engine asked about a brand by name cites its site almost always), answers replayed from the cache and engines that don't report what they read. A citation counts as a read. Your rate shows once 5 answers read your site on that engine, the other brands' once 5 brands from 3 other workspaces have 10 such answers each. Only these percentages are shared between workspaces: never a brand, prompt or site, and your own workspace is never part of the comparison.

Machine-translated Reddit threads

Reddit translates threads into other languages automatically and shows each translation at the thread's own address with a language parameter (?tl=de, ?tl=sl and so on). AI engines answering in your language cite these translations, but the discussion was written in another language, usually English. Asheard marks them wherever citations are listed (an answer's sources and the pages it read, Earned media and the citation gaps), with a link to the original thread, and Earned media and the action plan say how many answers cite them. To be part of that discussion, answer in the original thread.

Prompt stability

AI answers change from one check to the next, so one check is a sample, not a verdict. For each prompt and engine, Asheard counts how often you were named over the last 30 days. After 5 checks the result is labelled stable in (named in 80% of checks or more), stable out (20% or less) or coin-flip (in between). Hover a cell in the prompt × engine matrix, or click a prompt, to see it. Coin-flip prompts are where content work can tip the result, so Gaps and Fixes lists them. Answers reused from the 24-hour cache don't count as extra checks.

With several answers per check, each fresh answer counts as a check, so the labels come sooner.

Headline numbers and prompt intent

Every prompt has an intent: awareness (learning about a problem or category), consideration (looking for the best options), decision (prices, reviews, where to buy), comparison (X vs Y, alternatives to X) or branded (it names your brand). Asheard sets it when the prompt is added; change it on the prompt row if it is wrong.

A branded prompt names you, so the answer almost always does too, and a comparison prompt is about the brands it names. Both would inflate your numbers, so the headline numbers leave them out: visibility, its trend and per-engine split, win rate and share of voice, on the AI visibility page, Insights, reports and share links. The By intent card shows visibility for every intent, including the two left out, and Include them counts them in the headline again for the whole project.

Gaps and Fixes

Gaps and Fixes (Gaps in the sidebar) lists what to fix for a project from the last 30 days of answers: the action plan first, then the detail behind it: where your brand drops out of the answers, broken links, coin-flip prompts, citation gaps (pages cited in answers that name your competitors and not you; your competitors' own pages don't count), wrong facts, earned media and the searches where your site never came up. Share links include it as its own tab.

Action plan

The action plan turns what Gaps and Fixes finds, and your latest checks, into one list of actions, grouped by where the fix is: your own site, earned media (other sites) and technical. It reads the last 30 days of answers (negative answers, answers that name you but raise a concern, where your brand drops out, strengths the prompts asking who's best for them don't back up, coin-flip prompts, searches that never reach your site, other sites cited without you) and the latest fact check and broken links check, plus the buyer journey, page audits, recall and top-of-mind checks of the last 60 days. Each action says what to do, with the evidence behind it. Briefs only: the content is yours to write.

Each action has How to fix it: the steps, the change itself where the data allows (anything in [brackets] is yours to fill in, from facts you can check), how to check it worked, and a guide when there is one. Copy as prompt copies the action, its evidence, the steps and the change as one instruction for ChatGPT, Claude or any AI assistant, asking it to show every change before making it.

Priority is impact × ease, each from 1 to 3: 6 or more is high, 3 or more medium, the rest low. Impact is how many answers an action touches (10 or more is 3, 3 or more is 2) or, for the tool checks, how many engines; negative answers always count as 3, and a contradicted fact one step higher than its count. Ease is how much work the fix is: a redirect or a directory profile is easy, a new page or outreach takes more, press coverage and reference sites are hard.

Mark done keeps how the action's tracked prompts did in the 28 days before: how often they name you and cite your site. 14 and 28 days later the plan shows the same for days 1–14 and 15–28, from your scheduled checks, so there is nothing extra to run or pay for. A done action stays off the list for those 28 days, and comes back if the data still calls for it. Managers mark actions done and reopen them; share links show the plan and what was done, read only. New high-priority actions come in the AI visibility alert digest, each once.

Buyer journey check

Real buyers don't ask one question. The Buyer journey check (Tools → Buyer journey) has each engine answer five questions in one conversation, the way a buyer would ask them: the problem, the options, a comparison, the reviews and the final choice. You describe the buyer (who they are, their situation, what they need and their limits); that description opens the conversation, and the five questions come from a template in English, Slovenian, German or Croatian that you can edit.

For each engine you see which answers name you and at what rank, whether you are named at the start, where you drop out after being named, how many of the later answers still name you (persistence), and whether you are the one pick at the end, next to how often single checks of your tracked prompts name you. Every answer can be opened, with the competitors and other brands it names and what the engine searched for. ChatGPT, Gemini, Claude, Perplexity and Grok search the web as they answer; DeepSeek and Mistral answer from memory. Each turn costs the same as a normal check on that engine; if an engine stops answering, its remaining turns are refunded.

Answer-ready page audit

The page audit (Tools → Page audit) checks one page against one prompt: can AI engines reach it, quote it and use it as the answer? Enter the page and the prompt, or pick one of a project's tracked prompts. It is free and takes a few seconds; it reads the page, its robots.txt and llms.txt, and asks no AI model.

It checks: snippet controls (noindex, nosnippet, max-snippet and data-nosnippet for Google, plus noarchive and nocache for Bing, from the robots meta tags and the X-Robots-Tag header), whether the first 160 characters after the heading answer the prompt in its own words, the page's length against how much of a page Google's AI uses (about 540 words at most), whether the format fits the prompt (steps for a how-to, a definition for a what-is, a list for best-of prompts, a table for comparisons), robots.txt rules for this page grouped into search crawlers, crawlers fetching for a user and training crawlers, whether the text is readable without JavaScript, structured data and citability. The score is out of 100. Title, description, headings and llms.txt are shown for information.

The Snippet Lens, which shows the exact sentences an AI would lift from the page for the prompt, runs from the audit for its usual credits, once per audit, and is refunded when it can't read the page. Gaps and Fixes and the Google & Bing page link to the audit.

Broken links in AI answers

AI engines sometimes give links to pages of your site that don't exist: a made-up address, an old page or a subdomain you never had. The Broken links page (under AI visibility in each project) collects every link to your site from the AI answers of the last 30 days, written in the answer, cited as a source or among the pages the engine read, and checks each one on your server, one at a time. Tracking parameters such as utm_source are removed first.

A link is listed when the page is not found (404 or 410), when it sends visitors to the homepage, when it keeps redirecting, or when its subdomain doesn't exist. For the first two, the page suggests the closest real page from your sitemap (the Sitemap lines in robots.txt, else /sitemap.xml, /sitemap_index.xml or /wp-sitemap.xml), by the words the two addresses share, or else the section the link belongs to. Change any target, tick the links you want, and download 301 redirects for nginx, Apache (.htaccess) or the WordPress Redirection plugin (Tools → Import). Links the site refused to answer, or that answered with a server error, are listed separately: check them yourself. A page that says “not found” but answers as if it exists can't be detected.

The check is free. Managers start it; everyone who can see the project can read the result and download the redirects. Gaps and Fixes shows how many links don't work.

Earned media

The Earned media page (under AI visibility in each project) lists every third-party site the AI answers of the last 30 days cited for your prompts, by type: best-of lists, review sites, forums and communities, directories, news and media, video, reference, social and other sites. The type comes from the site (a review platform, a forum, a news site…) and, on news and other sites, from the page itself: a title or address such as “Top 10…”, “7 best…”, “…alternatives” or “Best … 2026” makes it a best-of list, /forum or /thread a forum, “review” a review. A site takes the type most of its citations have. Your own site and your competitors' own sites are left out, and so are answers replayed from the cache.

Each site shows how many answers cited it and on which engines, the competitors named in those answers, and whether you are named. The opportunity score is the number of answers citing the site × the competitors named there, less one when you are named there too (never below 0): a site the engines use a lot, where several competitors are and you aren't, scores highest. “Go after these first” lists up to 5 sites cited in 3 or more answers with an opportunity; Gaps and Fixes shows the top 3.

Best-of lists can be read, free: a manager starts it, and each list page the engines cited (up to 20, most cited first) is fetched once, like a browser would. The page's numbered headings (“1. …”, “#2 …”), or an ordered list that names you or a competitor, give each brand's place: “You're #4 of 10”. Without a numbered list, the page text says whether you are on it at all. What a list shows counts as being there for the score. Pages read in the last 7 days are kept when you read again, unless you changed your brand or competitor names; pages that couldn't be read are tried again.

Fact check

The Fact check page (under AI visibility in each project) compares your brand facts with what the tracked AI answers say about you. Add up to 8 facts, one per line: prices, what you sell, where you deliver, founding year, headquarters. Brand fact audit saves its facts to the project you pick there, and loads them back when you open it from this page.

The check reads the latest answer per prompt and engine that names you, from the last 30 days (up to 40 answers, newest first; answers replayed from the cache are left out). For every statement about you that touches one of your facts it gives a verdict: supports it, contradicts it (a different price, number, date, place, name or claim) or unclear (too vague to tell). Each one shows the answer's own words, the source the statement cites when there is one, and a link to the full answer. A quote is shown only when the answer says it word for word, or nearly, with every number the same. Facts an answer doesn't touch count as not mentioned.

It costs a few credits per batch of 8 answers, charged when you start it; a batch that returns nothing usable is refunded, and so is anything a check that stops early didn't reach. Answers already checked against the same facts keep their result and aren't paid for again; change a fact, or the brand name, and everything is checked afresh. Managers run the check; everyone who can see the project can read it. Gaps and Fixes lists the facts the answers contradict.

The Insights page

Insights pulls the measurements into one page per project: AVI scorecard and trend, a weekly narrative in plain language, the engine radar, the rank×AI quadrant and the analysis cards. What to do about them is on Gaps and Fixes.