Analytics & Trends

Overview

The Analytics and Trends pages transform your snapshot data into visual insights. While individual snapshots tell you what happened in a single moment, Analytics and Trends show you how your AI visibility is changing over days, weeks, and months.

AEO Optima provides two complementary views: the Analytics page for detailed breakdowns and the Trends page for high-level performance tracking.

Analytics Page

The Analytics page provides three primary visualizations that help you understand the shape of your AI visibility data.

Visibility Trend

A line chart that plots your brand visibility percentage over time. Each data point represents the percentage of snapshots that mentioned your brand on a given day.

Use this chart to:

• Identify upward or downward trends in brand mention frequency.
• Spot sudden drops that may indicate changes in AI model behavior.
• Correlate visibility changes with your content publishing or SEO activity.

LLM Comparison

Engines are no longer compared here on a figure blended across the four question types. The Platform comparison page (Performance › Platform comparison) compares every answer engine inside each question type, with one spread line per type; this section now shows those spreads in one line each and links to it.

Sentiment Distribution

A pie chart showing the breakdown of sentiment — positive, neutral, and negative — across the answers that both named your brand and were judged. An answer that never mentioned you has no opinion of you and is not counted, and an answer nothing judged is not "neutral". The CSV export carries the same three rows and, since 4 September 2026, the split per question type beside them; the Performance review shows that split on each question type's card.

Changed 4 September 2026. From that date an answer that does not discuss your brand carries no sentiment at all. Before it, such answers were stored as Neutral — the field could not hold "not judged", so an answer nobody had assessed was indistinguishable from one assessed as balanced. That change put two figures beside the chart: not judged (your brand was named, but no verdict was recorded) and did not name the brand (there was nothing to have an opinion about). Since 16 September 2026 the first of those has been replaced by one line beside the chart that accounts for every answer the tone split does not count, naming both how it was scored and what the judge found, with a count on each; did not name the brand keeps its own clause beside it. Nothing on that line is a slice of the pie, because a fourth slice would read as a fourth opinion.

This is a large change for projects whose answers rarely discuss them by name. A neutral share measured before 4 September 2026 is not comparable with one measured after it.

Changed 16 September 2026 — the mix stops counting stand-ins, and stops blending question types. Three fixes landed together.

First, the sentence above became true on every chart. Four tallies on the sentiment page — per engine, per question category, the trend, and the per-question breakdown — were still counting an unjudged answer as neutral, while the pie beside them correctly left it out. The per-question detail page did the same on all three of its charts. Every chart on both pages now counts tone the same way, so no two figures on one screen can divide by different populations again.

Second, the tone split now counts judge-scored answers only. When the sentiment judge cannot be reached, a word-list fallback still records a verdict so the capture is not lost — but a word list is a stand-in for a reading, not a reading, and it was being averaged in as though it were one. Those answers have not disappeared. Two separate things are recorded about every answer and an answer is usually in one of each, so the line under the chart reports both. How it was scored leads — judge-scored, word-list scored, or no record of how it was scored — and what the judge found is nested under it — accepted, named, nothing to judge (your brand was named, but the answer held no sentence evaluating it), judge unavailable, judge not run, or no judge outcome recorded — with a count on each, plus the answers that never named your brand. A group with no answers in it is not printed. Until 16 September 2026 each answer was forced onto one of those two, so an answer that was word-list scored and carried a judge outcome was counted only under the outcome: one project (Project A below) was told 38 of its answers were word-list scored when 75 of them were.

Third, the headline stopped being a share. Your four question types are four different games with four different win conditions, so a positive rate averaged across all of them describes none of them. The cards and the donut on Sentiment now print counts of judged answers — positive, neutral, negative — which may legitimately be totalled. Percentage shares are shown where they mean something: per question type, on the cards at the top of that page, and per question, on a question's own detail page. The trend and the per-engine views still put all four types together and now say so under the chart.

What this looks like on live projects, each with its window named, and each figure one a page actually prints. Projects are named by letter, and a letter is the same project in the table further down:

• Project A, 30 days to 16 September 2026 — 168 answers were judged: 43 positive, 118 neutral, 7 negative. Per question type: Brand · Informational 30.0% positive of 90 judged, Brand · Transactional 23.2% of 56, Non-brand · Informational 23.1% of 13, Non-brand · Transactional 0.0% of 9. Of the 243 answers that named the brand, 75 not counted — all word-list scored: 38 no judge outcome recorded · 37 named, nothing to judge.
• Project C, same window — 16 answers were judged: 8 positive, 8 neutral, 0 negative. Per question type: Brand · Informational 4 of 11 positive (36.4%), Brand · Transactional 4 of 5 (80.0%). 1 not counted — word-list scored, no judge outcome recorded.
• Project J captured nothing in the 30 days to 16 September 2026, so its figures are over 60 days to that date and are not comparable with the two above: 48 answers were judged: 37 positive, 11 neutral, 0 negative. Every one of them is Brand · Informational, which reads 77.1% positive of 48. 3 not counted — word-list scored, no judge outcome recorded.

A tone split measured before 16 September 2026 is not comparable with one measured after it.

Changed 17 September 2026 — the rest of the product caught up with the Sentiment page. The two changes are a day apart, and it matters which is which.

• The Sentiment page and the per-question detail page changed on 16 September 2026.
• The generated report, the exported PDF, the shopping view, the per-language analysis and the tools an AI assistant calls followed on 17 September 2026.

For that day in between, the same question asked on two screens had two answers: the page counted judge-scored answers and everything else still counted word-list verdicts as judged ones. If you pulled a report on 16 September, you are holding a copy from that gap — it counts the larger population, and it does not carry the basis line the copies from 17 September onward do.

The fix moves numbers on the later surfaces, downward, because they were counting more answers than they should have been.

What it costs, measured over the 30 days to 16 September 2026, on every project whose figure moves. "Counted before" is the answers those surfaces counted; "counted now" is the answers a judge scored:

Across the platform that is 1,875 answers counted before and 1,777 now — 98 answers, 5.2%. The nine rows above sum to 1,717 rather than 1,875 because the platform figure also counts the projects whose figures did not move at all.

The per-project figures come first because the average describes none of them: eight of the nine move by 8% or less, while Project A loses close to a third of its counted answers.

For Project A this is not a restatement. Its Sentiment page already showed 168 on 16 September. What changed on 17 September is that its report, its PDF, its shopping view, its per-language analysis and its assistant tools finally agree with the page.

Entity analysis did not move, and that is not an oversight. The related-entity tone on the Entity page reads the per-brand rows the sentiment judge itself writes, and has always asked the database for judge rows only. It was already counting what everything else now counts.

One tool still counts every brand-mentioning answer, and says so. The assistant tool get_dashboard_metrics returns a sentiment score and a polarity — averages of each answer's numeric score, not counts of verdicts — together with a positive rate, which is a count of verdicts. All three still cover every answer that named your brand, judged or not, exactly as their descriptions say.

The reason differs between them, and it is worth being exact because the earlier wording covered only two of the three. The score and the polarity are a different measurement read off a different column, and the score cannot express "nothing was judged" at all — it would have to report 0, which on its scale means the worst possible tone, so moving it needs a nullable replacement first. The positive rate has no such obstacle; it is simply not changed yet, and it is listed here so it is not mistaken for a figure that already moved. Both are their own dated change rather than a silent one folded into this.

What moves an answer out of the count. Two things, and only the first is what the table above is made of: a verdict a word list wrote when the judge could not be reached, and a row that does not record how it was scored at all. Over the 30 days these figures cover, every one of the 98 dropped answers was a word-list verdict and none was a missing-provenance row — so the explanation above is complete for that window. It is not complete platform-wide: older answers carrying no provenance key stop counting too, for the same reason, since silence is not a record that a judge did the reading.

The quick brand check never scored tone, and from 17 September 2026 it says so. The 3-model quick audit — the brand check on our public site, the one shown while you set a project up, and the assistant tool quick_audit — answers one question: does each engine name your brand. It does not run the sentiment judge. Until this date it reported a tone anyway. The assistant tool returned a fixed "neutral" for every answer that named the brand, the same value every time, whatever the answer said. The public check scored a seventeen-word list of its own — ten words counted as positive, seven as negative — which is neither the judge nor the method any other surface uses. Neither figure was a measurement. From 17 September sentiment is null on both, with an additive sentimentBasis of not_scored beside it saying why. The field is unchanged, so nothing reading it breaks — but its VALUE moved: code that treated a quick-audit sentiment as a verdict was reading an invention and now reads null. Tone is judge-scored on your tracked questions, and that is the figure to use.

A quick brand check now reports over the engines that answered, and from 17 September 2026 a refusal is never printed as a zero. The 3-model quick audit — the brand check on our public site, the one shown while you set a project up, and the assistant tool quick_audit — used to divide the mentions by the number of engines it TRIED. So when an engine could not be reached, it counted as an engine that had declined to name your brand. On 17 September 2026 our model provider's balance ran out and every one of the three calls came back refused; the check published { mentioned: 0, total: 3, percentage: 0 } — no AI engine mentions your brand — and cached that answer for 24 hours, serving it to everyone who asked about the same brand for the rest of the day. It was not a measurement. We had not managed to ask anybody. From this date: visibility carries answered and errored beside the existing mentioned and total; the percentage is computed over answered; percentage is null — never 0 — when no engine answered, with a machine-readable reason of engines_unreachable and a plain sentence in note. The field is unchanged, so nothing reading it breaks — but its VALUE moved, the same way sentiment moved earlier on this page: code that assumed percentage was always a number now has a null to handle, and that null means we could not ask, which is different information from a zero. A partially answered check still gives you a number, computed over the engines that did answer, and every screen prints "N of M engines answered" beside it. total keeps meaning what it always meant — the engines we tried — and there is deliberately no second field that repeats it. An audit in which nothing answered is no longer cached at all: a failed check is not a result to hand to the next visitor.

The quick brand check reports whether each engine NAMES your brand, and from 24 September 2026 it says that naming is not knowing. The 3-model quick audit — the brand check on our public site, the one shown while you set a project up, and the assistant tool quick_audit — asks each engine about your brand by name, and counts an engine when its answer contains the name. An engine that says it has never heard of you still writes your name in saying so: "I don't have reliable information about Zorvane" names Zorvane. So a brand nobody has heard of could be named by all three engines, and the check printed that as visibility. From this date it says what it measures. Each engine's line reads Named or Did not name your brand, and under it are the engine's opening lines — about 300 characters from the start of its answer, for every engine that answered, named or not — so you can read what it actually said, a refusal included. Until now the public check showed a short window starting just before your name, and the assistant tool the first 200 characters untrimmed, so an opening "I don't have reliable information about…" could be cut away. The headline reads "N of M engines named your brand in their answers" ("in its answer" when one engine answered), counted over the engines that answered, and the screens no longer draw a bar or print a percentage, because a naming rate drawn as a gauge reads as a score of how well AI knows you. percentage is unchanged in the API: it is the named rate, engines that named you divided by engines that answered, and it is still null, never 0, when none answered. The payload carries an additive namingNote with the sentence every screen prints: "Naming is not knowing: an engine that says it has never heard of you still names you. Inside AEO Optima, your non-branded questions — the ones that do not name you — show whether AI brings you up on its own." Checks cached before this date, or under an earlier wording of that sentence, are run again on the next request rather than served.

From 25 September 2026 every captured answer records why the engine stopped, and an answer that stops at its length limit is recorded as cut off; rates do not yet treat cut-off answers differently. On the model gateway an answer's output budget (max_tokens) covers the engine's reasoning and its visible text together, so a thinking engine under a budget sized for plain answers could stop mid-sentence; the gateway reports that as finish_reason: "length". Each answer now stores finish_reason, with the provider's own native_finish_reason beside it, read from the provider's response and never guessed from the text; answers captured before this date carry no such record, and nothing stored was re-scored. In the same change every engine's capture budget is one number, 4,096 output tokens (it was 2,048), with a bounded reasoning allowance of 1,024 tokens for the engines that honour one. Measured on 24 September 2026 in the capture's own shape, gemini-3.6-flash stopped on its own at 2,473 tokens, qwen3.7-plus at 2,475, kimi-k2.6 at 1,730 and deepseek-v4-pro at 1,335, each a whole answer; mistral-medium-3-5, which writes very long answers, still reached the limit at about 11,000 characters, and such answers are recorded as cut off. The quick brand check has its own rule, on its own surface. A reasoning engine — or one the registry has not described — is asked with 1,500 output tokens instead of 500 (gemini-3.6-flash spent 479 of 500 on reasoning and returned 65 characters of junk); an answer the engine cut off is marked cutOff: true on its model entry and counted as neither named nor not named, visibility.cutOff says how many, and the headline reads "2 of 2 engines named your brand in their answers; 1 engine's answer was cut off"; when every engine that answered was cut off, percentage is null with reason: "answers_cut_off", a different condition from engines_unreachable; and a check in which any engine was cut off is not cached, so the next visitor gets a fresh chance at a whole answer.

Use this chart to:

• Get a quick read on overall AI perception of your brand.
• Track whether the sentiment balance is shifting over time.
• Identify when negative sentiment spikes and investigate the cause.

Intelligence Scores

The Analytics page surfaces six proprietary Intelligence Scores — composite metrics that distill your snapshot data into actionable numbers. Each score is displayed as a KPI card with a 0-100 value and trend indicator.

These scores update automatically as new snapshots are captured. Together, they provide a comprehensive view of your AI brand health that goes beyond simple visibility percentages.

Tip: Use CMCS to identify models that describe your brand inconsistently, then investigate the specific prompts where discrepancies appear. A low CMCS often indicates opportunities to improve your content for specific AI engines.

For a deep dive into each score's methodology and how to improve them, see the dedicated Intelligence Scores page.

Share of Model

Share of Model is the platform's name for its core visibility measurement: the percentage of sampled AI runs (per prompt and engine) in which your brand appears. It is the same computation as Mention Rate — runs mentioning your brand divided by total sampled runs — presented under the name the AEO/GEO method uses for it.

The name matters because of how answer engines actually behave:

• AI answers are a stochastic distribution. Independent multi-run studies find the same prompt almost never returns an identical brand list twice, so a single answer or a "rank" is not a measurement.
• The distribution has structure. The consideration set is stable even when order is not — strong brands still appear in a majority of runs (often 55-80%). Inclusion frequency across many runs is therefore the honest unit of measurement.
• Sample size is part of the number. The measurement discipline targets 30-50 runs per prompt and engine. The per-prompt trends page flags any prompt-engine pair with a smaller sample: fewer than 10 runs is marked low confidence, and 10-29 runs is marked below target sample — treat those readings as directional until more runs accumulate.

You will see Share of Model on the Trends page (the headline KPI and the main chart) and per engine on the per-prompt trend detail, always with the run count alongside so a thin sample is never presented as truth.

How a brand mention is detected

An answer counts as mentioning your brand when your name appears in it as a whole word. Because AI answers rarely write a company's full legal name, the detector also matches shortened forms of it — "Northwind Steel India Limited" is recognised as "Northwind Steel", and a one-word form is used only when it is distinctive enough to stand alone. A short or ordinary word is never matched on its own: on a project tracking "Kira Rebar", a mention of "Kira Power" is not a mention of your brand.

A near-miss spelling of a longer name counts only when it carries the word that identifies you. Since 30 September 2026 a company that shares only the other words of your name is never read as a near-miss of it: on a project tracking "Northwind Healthcare", an answer about "Kira Healthcare" does not mention your brand, while a typo such as "Northwnd Healthcare" still does. Answers captured before the change keep the reading they were given; their algorithm version records which rule produced it.

A short form made from part of your name is counted on its own only when it is five letters or longer. On a project tracking "NWH Healthcare", an answer that writes only "NWH" is not a mention of your brand: "NWH" is part of the name and under five letters, and a short form like that is usually another company's name as well, so counting it would credit that company's mentions to you. (That name's acronym, "NH", is under three letters, so it is not recognised either.) A recognised acronym of three or more letters still counts: a project tracking "Northwind Kira Systems" is recognised by "NKS" written in capitals.

Since 5 November 2026 a one-word form that many companies share never counts on its own. "Shree", "Shri", "Sri", "Sree" and "Startup" are the list: on a project tracking "Shree Northwind", an answer about "Shree Kira Traders" does not mention your brand, while "Shree Northwind" itself still does. Another way of writing your own name counts once it is saved as one of your names. This list has a limit. Words alone cannot tell your own product from another company, so any other one-word form still counts whatever follows it: on a project tracking "Northwind Cables", both "Northwind Gum" and "Northwind Steel" count. Owners and admins settle those under Names that may be your brand on the Competitor comparison page, and saying a name is not yours corrects the stored answers.

Search-surface answers — Google AI Overviews and AI Mode — go through the same detector as chat answers since 8 September 2026. Before that date the search leg counted a mention only when the exact configured name appeared as a whole word, so a shortened form that counted on ChatGPT did not count on an AI Overview. Answers captured before the change keep the reading they were given; their algorithm version records which rule produced it.

Competitors are matched the same way, from the name and the website you configured, so a competitor entered as a web address is still recognised by the name an engine actually writes.

Correction applied 3 September 2026. Brand-name matching previously relied on a fixed list of company-name endings, so a name it could not shorten — anything ending in an unlisted word, or containing a typo — matched only in full. Answers that plainly named those brands were recorded as not mentioning them.

The detector no longer depends on that list, and the correction was applied to all captured history, not only to new captures. Affected projects will see their historical Share of Model rise; that step is the measurement being corrected, not a change in performance. Fixing it only going forward would have put a false improvement into every affected trend line on the day of the change.

Answers that came back empty are now excluded from these figures entirely. An empty answer is not evidence that your brand was absent, and counting it as such understated visibility on the engines that produce them.

Second correction, applied 4 September 2026 — what the first one got wrong. The 3 September detector could accept a shared word as a company name. A competitor called "Northwind Technologies" was recorded on answers that only contained the word "Technologies" (from a different company's name); a country's name, or the word for what a product is, could stand in for the brand; "International School" for a named school. On one project a competitor appeared on 141 answers whose text named it 19 times.

The detector now requires the identifying part of a name — never a sector word, a place, or a common word one letter away — and, for a company whose domain abbreviates its name, recognises the name people actually write. The correction was again applied to all captured history. What changed, in full: 178 brand mentions were removed (answers whose text does not contain the brand at all — 133 of them on one project where the word for its product had been read as the brand), 1 was added, and competitor attributions were corrected on 2,305 answers. Of the 1,013 brand mentions added on 3 September, at most 38 were reversed today; the rest stand. Answers from Google where no AI Overview was generated are never counted as a brand mention — the text of such a row is the ordinary search results, not an AI answer. From 9 October 2026 the same holds when an AI Overview exists: only its own text counts, never the sources it cites or the organic results below it (see SERP Tracking).

Every stored answer now records how each brand and competitor was matched and at what confidence, so a mistake of this kind can be found by a query rather than by reading answers by hand. Before today that record did not exist, which is why the size of the 3 September error could only be measured by re-running the correction.

The AI Surfaces split

The Analytics page also splits Share of Model by surface family — chat LLMs vs Google AI Overview vs Google AI Mode — so movement on one surface is never hidden inside a blended average. Each surface on the card discloses its capture method (native browser capture for the Google AI surfaces, provider APIs for chat LLMs). See SERP & AI-Surface Tracking for how the Google AI surfaces are captured.

Trends Page

The Trends page takes a higher-level view, designed for quick weekly reviews and executive-level reporting.

KPI Cards

Four key performance indicators are displayed at the top of the page:

Visibility Area Chart

A filled area chart showing brand visibility over the selected time range. The area fill makes it easy to see the overall trajectory at a glance, while the line edge shows daily fluctuations.

LLM Performance Bars

The per-model bars are gone from Trends: engines are compared inside each question type on the Platform comparison page, and this slot shows those spreads in one line each, with a link to it.

Top 5 and Bottom 5 Prompts

Two ranked lists that highlight your best-performing and worst-performing prompts within the selected date range:

• Top 5 Prompts: The prompts with the highest brand visibility. These represent your strongest areas of AI presence.
• Bottom 5 Prompts: The prompts with the lowest brand visibility. These represent your biggest opportunities for improvement.

Tip: Check the Bottom 5 Prompts list weekly. These are your biggest improvement opportunities. Focus your content strategy on the topics where AI models are not yet mentioning your brand.

How rates are counted (from 4 September 2026)

Four rules apply to every rate on the Performance review page and the home page's pool cards, and they are being extended to the rest of the product.

Four kinds of question, never added together. Every tracked question sits in exactly one of four pools — whether the asker already named a company (brand / non-brand) and whether they are choosing or learning (transactional / informational). A question is informational when its intent is Informational or Hybrid; every other intent is transactional. From 23 Oct 2026 hybrid questions (information plus buying criteria) count as informational, in every period shown: a non-brand one with the informational questions, not the transactional ones, and a brand one with the brand informational questions. Reports, digests and alerts before that date counted them as transactional. Each pool has its own headline measure: named-in-answer rate for non-brand transactional, cited rate for non-brand informational, own-source share and facts stated correctly for brand informational, sentiment and accuracy for brand transactional. There is no combined score across pools anywhere in the product, because a question about you and a question about your category are different contests.

Rates count questions, not answers. A question is asked many times per cycle — several runs on each engine — and answer engines are stochastic. Mention, citation and top-three rates therefore divide questions by questions: a question counts toward the numerator when it qualifies in the majority of its runs in the window. "Total mentions" is the one figure that stays at answer level; it is a count, shown beside the rate and never as a percentage, because it can exceed the number of questions.

The denominator is the contest you can enter. Non-brand rates divide by the pool's contestable questions — those where at least one company was named or one source cited. A question no engine answers with any company is open ground, reported as its own figure, never as a loss. Brand-pool questions name you by construction, so their rates divide by the pool's tracked questions.

Movement smaller than the noise floor is not reported as change. The floor is fixed: a share metric has to move more than 6 points (4 points for sentiment) before the product calls it a change, however many answers sit behind it, and projects differ in how many answers they collect per engine per cycle. Smaller moves are labelled as inside the noise floor. A single question's change between the last two complete cycles ("What moved" on Home) is listed only when each of the two cycles holds at least 10 answers on that question; with fewer, a question can only move in large steps (on three answers, by thirds), so it is counted ("n questions changed, but on too few answers a cycle (under 10) to call it a change") and never given a point change (from 22 October 2026). A trend direction needs at least four captured cycles; coverage is judged against the captures your schedule actually promised, not against calendar days.

An answer nobody judged is never counted as neutral. From 4 September 2026 an answer that does not discuss your brand carries no sentiment; before that date such answers were stored as Neutral. Every sentiment share divides by the answers that named your brand and carry a verdict a judge produced, and what is left over is reported beside the shares rather than folded into them: one line accounting for every answer the split does not count — how it was scored, with what the judge found nested under it — and, separately, the answers that did not name your brand. The Sentiment page adopted that rule on 16 September 2026; the generated report, the PDF export, the shopping view, the per-language analysis and the assistant tools followed on 17 September 2026, and from that date all of them count the same answers. Each report and PDF from 17 September onward carries the basis in print, beside the figures, so a copy you keep says what it counted.

Tone is reported per question type, over the answers that judged you. Each question type's card on the Performance review shows positive / neutral / negative shares over the answers that both named your brand and carry a judged verdict — never over every answer. Below four judged answers the card shows the count and says the sample is too small; three shares drawn from two answers look exactly like three shares from two hundred, so the picture is what is withheld.

A question type's trend is plotted per capture cycle, on your own cadence. The "named in the answer, per capture cycle" line on each card has one point per capture cycle — a week for a weekly project, a month for a monthly one — and each point is the share of that type's questions captured that cycle where you were named in the majority of runs. A cycle with no captures is a gap, never a zero, and the line does not appear until two cycles have been captured.

Engines are compared inside one question type. Since 26 October 2026 the Platform comparison page has three tabs, one per question type, each with its own measures in a fixed order:

• Non-brand · P1: mention rate, top-3 presence and positive. Citation is read on informational questions (P2 + P3), as on the Overview, and a line under the table says so.
• Brand · P0: accuracy, positive, neutral, negative, competitors appearing and own-source citations. Own-source citations is read over the informational brand questions, as on the Performance review: of those answers that cited anything, how many cited your site.
• Non-brand · P2 + P3: mention rate and citation rate.

A measure we do not take yet is not shown as an empty column; one line under the table says so ("Critical is not measured yet.").

By engine shows every engine for the last complete month:

• All engines (pooled) comes first. It is every engine's answers together, the same figure the Overview prints.
• Each engine's best value in a column is highlighted. The highlight marks the best of these engines; it is not a judgement.
• Google's AI Overview and AI Mode, and Bing's Copilot answer, say how many of the tracked questions they showed an answer for ("12 of 20 shown"). Their rates are over the answers they showed.
• Accuracy checks also run on AI Overview and AI Mode answers, and each surface shows its own Accuracy where the checks name it. Checks that arrive without their surface are counted in the pooled row only; in such a month those rows say so, and the pooled row's hover says how many of its checked answers were search answers.
• One line per tab gives the spread on the tab's first measure, in points. It keeps both floors: an engine with fewer than five contestable questions is left out of it, and a spread under the six-point noise floor is reported as inside the noise floor.

The weighted row is a presentation, not a measured rate. Where a dated public figure for each engine's share of assistant use exists for the project's home market, a row blends the engine rows by that share.

A project has one home market. With one active country, it is that country. With more than one, an owner or admin chooses it on Locations; until then there is no weighted row and the page says so. Only India has sourced figures today, so any other home market shows no weighted row. Where a project captures in other countries too, the row blends the answers from all of them, weighted by the home market's share of use, and the page names the others as having no weighted row of their own.

• For India this is StatCounter's AI chatbot market share, which counts clicks from assistants to websites, so assistants that link out weigh more.
• Search surfaces, engines with no published figure, and engines with too few answers to judge are left out and named.
• Counts are never weighted, and the row never mixes question types.
• Nothing is judged against it, and no alert or report reads it.

By month shows one engine, or all engines pooled, month by month under quarter headings. Each row is shaded against its own complete months, with the strongest shade on that row's best month. The shading is never a pass or a fail, and the figure on every shade keeps at least a 4.5:1 contrast in the light and dark themes.

The Volume covered row is your team's own figure for the share of search demand the tab's questions cover, entered on Figures you enter and read by the same rule as the Overview: a month without a figure stays blank, a quarter shows a figure only when every month of it has one, and on the P2 + P3 tab the P3 figure is shown with P2 named beside it. An entered figure carries the pen mark and is never shaded or judged. It is the project's figure, so it is the same whichever engine you select.

Each engine says what its answers search. Hover an engine's name on Platform comparison or the Performance review. Every chat engine we ask except Perplexity answers through one shared web search, so every model searches the web the same way for a question. That has been so since 6 Aug 2026; earlier answers used no web search. Perplexity searches its own index. Google AI Overview and AI Mode answer from the Google index, and Bing Copilot from the Bing index. For ChatGPT, Gemini, Claude and Grok the hover also gives the index that company's own app searches, dated, because that is about their app and not about how we ask. Generated reports print the same label under each engine.

Each question type carries its code. P0 is either brand type. P1 is non-brand transactional. P2 + P3 is non-brand informational, because hybrid questions (information plus buying criteria) count with the informational ones.

Segment Filter

Both the Analytics and Trends pages include a Segment Toggle — the segment pills of the lens filter bar, which also offers Intent, Journey, Topic, Tag, Group, and Orbit filters — that lets you filter all data by prompt type:

The Non-Branded view is particularly valuable because it shows your true organic AI visibility — how often AI models recommend you when users aren't specifically asking about you. The Competitor view reveals how you perform in direct comparison queries.

The segment toggle is URL-synced (?segment=branded), so filtered views are bookmarkable and can be shared with team members.

Tip: Compare your Branded visibility (should be high) against your Non-Branded visibility (the real signal). A large gap means AI models know your brand but don't organically recommend it — this is your biggest opportunity for improvement.

Date Range Filters

Both the Analytics and Trends pages support date range filtering to focus your analysis on the time period that matters most.

Select a date range using the filter controls at the top of either page. All charts and KPIs update immediately to reflect the selected period.

How to Use Analytics Effectively

Weekly Review Workflow

1. Open the Trends page with a 7-day filter.
2. Check the Trend Direction KPI — is visibility moving up or down?
3. Review the Bottom 5 Prompts — are there any new entries since last week?
4. Switch to the Analytics page for deeper investigation if you see unexpected changes.

Monthly Review Workflow

1. Open the Trends page with a 30-day filter.
2. Compare Avg Visibility and Avg Rank to the prior month's values.
3. Check the LLM Performance Bars — has any model's behavior changed significantly?
4. Review the Sentiment Distribution on the Analytics page — is the positive/negative balance stable?

After a Content Campaign

1. Set the date range to cover the period before and after your content was published.
2. Watch the Visibility Trend line chart for inflection points.
3. Check the LLM Comparison to see if the new content improved visibility across all models or just specific ones.

Analytics and Other Features