How to Read Your Five-Engine AI Visibility Report

This guide walks through the five-engine report section by section: what each part measures, why it answers a question your analytics and rank tracker cannot, and the specific action it points to. It assumes you have the report open. If you are reading the free single-engine report instead, how to read your AI visibility report covers that one.

What is in your five-engine report, and what order should you read it in?

Your five-engine report is three stacked layers: a headline score at the top, a cross-engine comparison in the middle, and one engine's full breakdown underneath. The headline score is the part most people misread: it is a single engine's figure, not an average of five. Skavora promotes the first engine that returned a healthy result to the top of the report, and in practice that is Gemini, which sits first in the engine order. An engine that degrades loses its own score rather than dragging yours down, so if Gemini degrades the next healthy engine is promoted in its place, and the label above the single-engine breakdown names whichever engine it is. Each engine's score also sits in its own benchmark card further down. The headline repeats the promoted engine's card, and the other four can differ widely from it.

On screen the order is fixed. The header carries your brand, category, country, scan date, a badge saying how many engines returned a result, and the PDF download. Amber notices come next: one for every engine that was skipped or failed, plus notices where an engine was measured on a sample of the question set or some queries came back empty. Then the hero score, the branded and non-branded coverage split, and a panel confirming the report is saved against your checkout email address. If the brand could not be confirmed against a same-name business, the hero shows an amber card and no score at all. If no engine returns a trustworthy result there is no report to read at all, and your purchase stays valid. The cross-engine section follows, then the single-engine deep dive, the recommendations, and the follow-up question box.

The cross-engine section is the middle layer. It runs five cards in this order: a benchmark card per engine (score out of 100, band, non-branded coverage, share of voice, average position, or a "Not measured" or "No result" state when the engine did not run), a grid of every buyer question against every engine that scored, the exact pages cited across engines, the sources each engine trusts, and the buyer questions where no measured engine named you. A "Visibility trends" card closes the section: it is a placeholder and charts nothing today. Two cards drop out when the data is thin. The pages card needs page-level citations, and the sources card needs at least two engines that returned citations, so some reports will not show them.

That structure answers something a rank tracker cannot reach: whether your absence is one engine's opinion or every engine that measured the question agreeing. Rank trackers measure positions on a results page, and analytics only describe visits that already happened. A single score hides all of it.

Read the first pass out of order, cross-engine before per-engine:

1. The engine benchmark cards, the row under the "across N AI engines" heading, for the scores side by side and which engines did not run.
2. "Where you're invisible across every engine", the questions no measured engine answered with your name, up to twelve of them.
3. "The exact pages AI cites", starting with the pages more than one engine cites.
4. "Where you appear, engine by engine", top rows first: it is sorted with your biggest gaps at the top.
5. The single-engine deep dive last, for competitor rankings, sentiment and source detail.

Signal and noise separate cleanly here. A gap in one engine is weak evidence and can reverse on a later run; a gap in every engine that measured the question is a content problem you own. Read the grid markers precisely: a dash is a measured absence, "n/a" means that engine returned no answer at all for the question, and a small dot means the question was never measured for that engine. Only the dot is missing data. Start with the benchmark cards, then the cross-engine gaps, and leave the headline score alone until you have read both.

How do you read the per-engine benchmark?

The per-engine benchmark is the grid of cards at the top of the multi-engine section, one card per engine the scan covered, in a fixed order: Gemini, ChatGPT, Claude, AI Overviews, Google AI Mode. Each card carries that engine's own score out of 100 and its band (Dominant from 80, Strong from 60, Emerging from 40, Low from 20, Invisible below that), then three figures: non-branded coverage as a ratio, share of voice as a percentage, and average position, labelled "Avg position". Every figure comes from that engine's own answers. Nothing is pooled across engines, and the heading above the cards counts only the engines that produced a score.

A card without a score is never a zero.

A degraded engine carries no metrics at all. When fewer than half the questions expected of an engine come back measurable, or answers arrive that nothing could score, the card shows no score rather than a low one, because a number built on that base reads as poor visibility when it is really a poor scan. Those cards show "No result", sometimes with the partial badge, and with a reason line underneath when one was recorded. Treat them as missing evidence rather than bad news.

Average position shows a dash until at least three non-branded answers placed your brand in a ranked list, because an average over one or two mentions is noise, and a position in a branded question means nothing. That has a cost: position carries 15 of the 100 points, and a withheld position scores zero there, so an engine with two ranked mentions cannot score above 85. The two Google engines measure a deterministic sample of the questions because each lookup is billed, so their denominators are smaller. Those denominators count only the questions that produced a trustworthy finding, and a question where the engine returned no answer at all, Google showing no AI Overview for instance, is one of them: it counts as an absence, not a failure. Compare ratios, not raw counts.

A wide gap between two engine cards is a statement about sources, not about your brand. Every engine gets the same questions in the same order, and no engine is told which brand the scan is measuring, so what differs is the material each one retrieved and trusted. A rank tracker cannot show you that, because it measures one index of pages, and analytics cannot either, because it only describes buyers who already arrived.

Treat a few points between engines as noise: answers are generated live and move between runs, and a re-run is a separate report. Treat a full band of difference, or non-branded ratios that differ by more than half, as real. Work in this order: the questions where no engine named you first, then the single engine with the lowest non-branded ratio among the ones your buyers actually use. Open that engine's list in "The sources each engine trusts", which appears once at least two engines returned citations, and pitch the domains it cites most, starting with any a second engine cites too.

How do I read the question-by-engine grid?

The question-by-engine grid is part of the full multi-engine report, headed "Where you appear, engine by engine". It puts one buyer question on every row and one AI engine in every column, and each cell records what that engine's answer did with your brand. Only engines that returned a trustworthy score get a column, so an engine that failed across the whole scan is missing from the grid rather than shown as a column of blanks. A four-column grid in a five-engine report is normal, not a fault. Every row repeats the question in full and tags it branded or non-branded.

Hovering over an n/a or a dot brings up a short tooltip. The n/a tooltip says the engine returned no answer for that question. The dot tooltip says only that the question was not measured for that engine, so it will not tell you whether the question was sampled out or whether something failed.

AI Overviews and Google AI Mode measure a deterministic, evenly spaced sample of the 25 questions, 15 of them by default, because every lookup on those two engines is a billed search. Unsampled questions appear as dots, are excluded from every figure in the report, and are never counted as absences. A banner near the top of the report names the engine and how many of the questions it measured. The sampler is deterministic in position: it picks the same slots in the question set on every run, so nothing about the choice is random. It does not promise identical question text next time, because the question set itself is generated fresh from live research.

A rank tracker tells you where a page sits for a keyword on one results page, and analytics only describes people who already arrived. The grid answers what neither can: for this specific buyer question, did the engine name you, name somebody else, or return nothing. It also separates a universal gap from an engine-specific one. Absence in every engine that measured the question points at the underlying material. Absence in one engine while the others name you points at what that engine draws on.

Rows sort by how many engines mentioned you, fewest first, with the original question order breaking ties. The sort key counts mentions only, not measurements, so a row of nothing but dots can sit at the top beside a genuine gap across every column. Read the cells before you read the position. The same caution applies to the header figure for questions where you appear across every engine that measured them, because a question only one engine measured still qualifies for that count.

Work down the grid to the first three non-branded rows carrying a grey dash in three or more columns, and skip any row that reached the top on dots. Those three are structural: the material an engine would need in order to name you either does not exist or is not in a form it can use. One grey dash beside four ticks is closer to run-to-run variance and does not justify a project. Write down the question text of the three you choose, because a re-run is a separate report and like-for-like comparison means comparing the same questions.

How do I use the list of exact pages AI cites?

The exact pages AI cites is a ranked list of the specific pages that ChatGPT, Claude, AI Overviews and Google AI Mode drew on while answering the buyer questions in this scan. It belongs to the multi-engine report, and only engines that returned a trustworthy result feed it. Each row carries the page address as a live link, with the https:// prefix dropped from the text, a badge for every engine that cited that page, the page's share of all page-level citations in the scan, and a count. Counting is once per page, per answer, per engine: a page quoted four times inside one answer counts once, and the same page cited by three engines in three answers counts three. Rows are ranked by count, highest first, up to twelve.

Addresses are normalised before counting. A leading www, any trailing slash and everything after the path, meaning query strings and anchors, are stripped, and the link is rebuilt as https. Variants of one page therefore collapse into a single row, which is what you want for a tracking parameter and not what you want on a site that identifies pages by query string, where several YouTube videos can land on one youtube.com/watch row and the link will not reopen the video. Open the link before you act on the row.

Only URLs with a real path appear, so Gemini never carries a badge here. Its grounding API reports the site it drew from rather than the page, because the page sits behind a Google redirect, so Gemini's citations arrive as bare domains and a page-level table filters them out. That is why the free Gemini report ranks sources instead, under the heading "Citation share by source". If no measured engine returned a page-level citation, this card does not render at all, which usually means the engines answered from their own knowledge or named only sites. Engine badges are hidden on small screens and the "you appear" marker on anything narrower than a tablet, so read this card on a desktop.

"You appear" is an answer-level marker, not a page check. It means your brand was named in at least one answer that cited that page. Skavora never fetches or crawls the page, so the marker does not confirm your brand is written on it, and its absence does not confirm you are missing from it. Treat it as a lead to verify rather than a verdict. Nothing marks your own domain or a competitor's, either, so those you spot by reading the addresses.

A rank tracker reports positions for keywords on a results page, and analytics reports the visits that arrived. Neither shows the pages an AI engine read while composing an answer: the answer often settles the question without a click, and when someone does click through to you, the referrer is the engine rather than the page it read. What this list adds is the reference material for your category at URL level rather than domain level. A domain list tells you Reddit matters. This tells you which thread.

Work the list in this order:

1. Pages badged by two or more engines. A shared reference pays back across the market rather than one engine's habit.
2. Pages on your own domain. Nothing marks them, so find them by reading the addresses. They already earn citations, so keep them current and extend the sections that answer the questions you are missing in "Where you appear, engine by engine".
3. Third-party roundups, reviews and guides where you are absent. Pitch inclusion to the author or editor named on the page, leading with a factual correction where the page is out of date.
4. Forum and community threads. Contribute under your own name; anything that reads as astroturfing costs more than it earns.

A count of one is a single answer to a single question, not a pattern, and every share figure describes this scan alone. Nothing here is tracked between scans, so a change is only visible if you run a second report and compare the two yourself. A row at one citation with one engine badge is noise unless the page is obviously important to your category. Two or more engines, or three or more citations, is a fixture worth the effort. Start with the highest-ranked page that is neither yours nor a competitor's, open it, and confirm whether your brand is actually on it before you write to anyone.

Each engine's source list is a separate outreach list, and the overlap figure tells you how separate

The card headed "The sources each engine trusts", in the multi-engine section of the paid report, gives each engine that cited something its own ranked list of cited domains, and states how many of those domains were cited by a single engine. The line above the columns gives four figures: the number of measured engines that returned citations, the total distinct domains across their lists, the count and percentage cited by one engine only, and the count cited by more than one. Each column holds up to six domains with a number beside each. That number counts answers, not citations. A domain quoted three times inside one answer still scores one, so the figure can never exceed the answers that engine produced.

Order within a column is your own domain first if it was cited at all, then the rest by citation count, so the top row is not always the most-cited site. Each engine records at most six sources per answer, and at most twenty domains per engine reach the report, so these lists describe what an engine leans on rather than every URL it touched. Gemini reports the site rather than the page, so its citations appear here at domain level only, which is also why Gemini is absent from the page-level citation card above.

The overlap figure answers something neither a rank tracker nor your analytics can reach: whether one body of source material shapes how AI assistants answer your category, or several separate ones do. Rank trackers report positions in a list of links, and analytics reports people who have already arrived. Neither shows which pages an assistant read before composing an answer that produced no click at all. A high single-engine share means a separate outreach list per engine rather than one, and a placement that moves ChatGPT may register nowhere in Google AI Mode. The figure is measured on your own category rather than assumed from a benchmark.

The cross-engine card shows domain and count only. Source types (Reddit, Review & comparison, Editorial & press, Competitor site and the rest) live in the "Where AI gets its answers" wheel and the outreach target list further down, both computed from the single engine named in the "in detail" strip rather than across every measured engine.

Treat a domain as a real target when it recurs, meaning a count of two or more, or when more than one engine cited it. Compare counts down a column, never across columns. Google AI Overviews and Google AI Mode measure an evenly spaced sample of the question set (fifteen of twenty-five by default) rather than every question, so their counts come off a smaller base and one citation there is thin evidence. Build a single list from the shared block, then add the recurring domains from the engine your buyers actually use. This card does not show whether you are already cited on a domain, so use the outreach target list below, which carries a presence column for the engine it measures. A domain that survives a re-run, which is a separate report, is a fixture worth pitching.

What should I do about the buyer questions where no AI engine names my brand?

Skavora's cross-engine gap list, headed Where you're invisible across every engine, is part of the five-engine report and shows the buyer questions where every engine that reached a verdict answered without naming your brand. A question qualifies only when each engine that produced a trustworthy score in your scan (Gemini, ChatGPT, Claude, AI Overviews and Google AI Mode) reached a definitive result for it, at least one of those engines answered and left you out, and none of them named you. The card shows the first twelve qualifying rows and does not state how many were found in total. Non-branded questions come first, and within each group the questions absent in the most engines come first.

Two states behave in ways worth knowing. An engine that returned no answer at all, such as Google showing no AI Overview for that query, counts as a measured negative rather than a hole in the data, so it does not disqualify a question, but it never appears in the absent-in line, which is why a five-engine scan can show a row naming only four engines. A question that any engine skipped or failed on is excluded entirely: the two Google engines measure a deterministic sample of the question set by default, 15 of the 25 questions, and a question they were never asked is not evidence of an absence. If nothing qualifies, the card does not appear at all.

This list answers a question your analytics cannot reach: which buyer questions produce an answer that leaves you out, and often names someone else instead. A rank tracker measures positions in a list of blue links, and an AI answer a buyer never clicks through leaves no session and no referrer in your reporting. A row here is also stronger evidence than a single-engine absence. Three of the five engines are Google products, but ChatGPT and Claude come from different companies drawing on their own source sets, so a unanimous absence is harder to put down to one model's judgement. The report measures the absence, not its cause, so treat the row as the finding and the cause as the thing to go and check.

Start at the top of the non-branded rows and work down. Read the constraint in each question, the budget, the location, the use case or the situation, then check whether any page you own answers that exact question in those terms. Then open "The exact pages AI cites" in the same report and look for pages cited often with no "you appear" flag: those are pages the engines drew on in this scan where your brand went unnamed, which makes being listed, reviewed or quoted on them worth pursuing. That card aggregates across the whole scan rather than per question, so read it as an engine's general reading list, not proof of what it read for this row.

Treat "named instead" as a pointer rather than a verified competitor. It is the most frequently named brand across those answers, taken as written, with no merging of spellings and no filtering of domains or your own aliases, so a marketplace, a publisher or a product line can occasionally sit there. The card gives one name, no count and no per-engine breakdown, so you cannot judge from it whether the engines agreed on that rival; check the name against "Who AI recommends most", which is computed from the one engine named in the "in detail" strip. One scan is one sample of a moving system, and a re-run produces a separate report with no automatic comparison, so build your plan around the rows that survive a second scan.

Can I ask questions about my report and re-run it with my own wording?

A paid Skavora report answers up to four follow-up questions about your own results, and it lets you rewrite the question set and put your own wording back through every engine at no further charge. The panel headed "Ask about your results" sits at the end of the report and is handed one scan's data as its only evidence: your headline metrics, every competitor row with its mentions, coverage, average position and share of voice, your ten most-cited sources, the source-type and social breakdowns, up to fifteen questions answered without naming you, and the recommendations. Answers quote those numbers back at you. A counter under the box shows how many of the four remain, and once they are spent the panel asks you to run a fresh scan.

The follow-up sees that one scan and nothing else. What it reads is the single-engine detail section, not the engine-by-engine comparison above it, so how the five engines differ is outside its evidence. Nor are there earlier scans in it, or trends over time, or traffic and ranking data; asked about any of that, it says so in a sentence and then answers from what was measured. It does carry your earlier questions, so each answer can build on the last. The four belong to the report in front of you rather than to your account, so a re-run or a reopened copy starts a fresh four, with an hourly allowance on the server behind that. Spend them on sequencing: which gap first, which source first, what to publish first.

The bar above the prompt-by-prompt table opens your question set for editing, seeded with the questions this scan ran. Reword them, delete the ones your buyers would never ask, and add your own, up to 25 questions of 300 characters each. Two measurement rules survive your edits: a question naming your brand counts as branded, and a question naming a platform such as Reddit, Trustpilot, YouTube or Google is flagged in amber, because the answer will lean towards the platform named. Re-running puts the edited set to Gemini, ChatGPT, Claude, Google AI Overviews and Google AI Mode, the two Google engines sampling it as they did before, and replaces the report on screen. A set you wrote yourself is also the surest way to make two scans comparable, since a generated set is researched again later and can come back different.

Download PDF opens your browser's print dialogue, where Save as PDF writes the report to a file. Skavora prints the page you are looking at rather than building a separate document, so the file and the report cannot disagree; the site navigation, the footer, the two buttons at the top and a few screen-only links drop out, and the cards, tables and figures print whole. Only one competitor row opens at a time, so open the row whose evidence you need, and ask your follow-up questions, before you print. Every paid report is saved against the email address used at checkout and reopens at /my-reports through a sign-in link. Follow-up Q&A works on a reopened report. The question editor does not, because a reopened report carries no proof of purchase.

Two things are worth knowing before you close the tab. A re-run does not overwrite the saved copy: the stored report is the first complete run of that purchase, so download the PDF of a re-run yourself if you want to keep it. Nothing in the report tracks your visibility over time either. The "Visibility trends" card at the end of the multi-engine section is a placeholder, an icon and two sentences, and it never fills with data however many reports you run; a second report weeks later is a separate purchase, compared by you. Spend your four questions now, re-run once if the set misreads your buyers, then save the PDF with the scan date in the filename.

What should I do in my first week with the report?

A first week with the five-engine report runs in three parts: read it on day one, pitch and publish across week one, and leave anything needing two scans until you re-measure. Day one is reading, not doing. Open "Where you're invisible across every engine", which collects up to twelve buyer questions that no measured engine answered with your name, non-branded first, each carrying the rival named most often in your place where one was found. Note the first three non-branded rows. Then record the four figures on each measured engine card (score, non-branded coverage, share of voice, average position) as the per-engine baseline you will judge the next scan against.

Two checks stop day one producing a bad list. In the grid headed "Where you appear, engine by engine", rows sort by how many engines named your brand and nothing else, so a row where nothing was measured can sit at the top beside a genuine gap: confirm a top row carries at least one grey dash before you act on it. On the benchmark cards, a dashed card reading "Not measured" or "No result" is an engine that did not report, not an engine that scored you nothing, so keep those out of the baseline. Then save the PDF; the report itself reopens from /my-reports, but only the file can be sent on.

The page-level list answers something analytics cannot: which pages the engines cited when they composed those answers, including the answers that named somebody else. Analytics records only the people who arrived, and an answer that omits you sends nobody, so the gap leaves no trace there. Rank tracking measures positions on a search results page, and the pages AI engines cite are frequently not those pages, as the 12% overlap figure earlier in this guide suggests. "The exact pages AI cites" gives you the addresses, badged with the engines that cited each one. If no measured engine returned a page-level address the card does not appear, and each engine's source list is your outreach list instead.

Week one is outreach first and publishing second, because outreach has the longer lead time. Work "The exact pages AI cites" from the top: start with the page carrying the most engine badges and no "you appear" flag, a priority-1 row, since one placement can pay out across several engines. Open each page and read it before pitching, because that flag describes the answers citing the page rather than the page itself. Then publish one page for each of the three cross-engine gaps you noted, using the row's own wording as the heading. Spend your four follow-up questions on sequencing, though those keep working on a reopened report, so they are not the urgent piece.

The question editor is the piece that will not wait, because it works only on a live paid report. If the generated questions miss how your buyers phrase the decision, edit the set and re-run it now, then keep your own copy of the wording, since nothing stores it for the next report. That copy is what makes re-measuring comparable, because questions are generated fresh each scan and a later run may not ask the same ones. Leave the comparison itself for several weeks, and make it yourself: nothing here watches for change. Signal is a question moving from grey dash to green tick in more than one engine, or a non-branded coverage ratio improving on the engine you worked. Noise is a few points of score, a position shifting by one, a single answer flipping, or a domain cited once. Chase what repeats across engines.