Working notes on measuring and improving AI visibility. Definitions with the arithmetic attached, failure modes we hit in our own audits, and no throat-clearing. Our own product flags that as an issue, so we do not get to do it here.
Prompts and surfaces need changing, and every change breaks a trend line. A request form, six change types, and the bridge period that keeps two versions comparable.
An AI search programme touches content, engineering, analytics and whoever reports upward. A RACI for its ten recurring jobs, and the two that fall between roles.
Model and product changes move AI visibility numbers without anyone touching your site. The sources to watch, the log fields, and how to annotate a series honestly.
The monthly pass that keeps AI visibility numbers trustworthy: data-quality checks first, exports second, numbers last. Twelve checks in order, plus a log template.
Four weeks, in order: sources and reachability, the instrument, two baseline windows, then an honest first report and one experiment. What each week ships and skips.
When an assistant gets your pricing, features or category wrong: log the claim, trace the source, fix what you control, re-test. The tracker fields and four routes.
Before calling a move in AI visibility real, know how much the rate moves when nothing changed. A worksheet to compute it from your own runs, with a worked example.
For each tracked prompt, name the page meant to answer it, then compare with what assistants cite. Five states, a mapping sheet, and the work each state points to.
Sort every URL an assistant cites into one of nine types, with written rules for the ambiguous ones, and see which sources the answers in your category lean on.
'We checked ChatGPT' can mean an app, an API, a plan tier, search on or off. The register that records exactly what you sampled, so two periods stay comparable.
A cropped screenshot of an AI answer proves very little. The fields to capture, a file-naming rule, and the context a screenshot needs before it goes in a report.
'Recommended by ChatGPT' is an advertising claim about a system that answers differently each time. The evidence it needs, and wording that survives a check.
A single trend line hides the run count, the spread and the model change. Seven chart rules, a spec to copy, and an illustrative before-and-after of the same data.
Every AI visibility number needs a methods note: prompts, assistants, runs, dates, exclusions and sources. A template to paste under every report, field by field.
Google now counts AI Overview and AI Mode impressions; Bing counts Copilot citations. What each report measures, and a layout that never adds them together.
Experiments in this channel rarely produce a clean result, so they run indefinitely. Four stopping rules, decided in advance, including the one that admits nothing was learned.
AEO programmes depend on context that lives in one person's head. The nine artefacts a successor needs, and the one whose absence makes the history worthless.
Most AEO reviews present activity and call it progress. Six agenda items, in an order that makes an honest review possible and a flattering one hard.
AEO spend is easy to justify with a scary framing and hard to justify with evidence. What to fund, what to defer, and the number nobody can give you.
Your competitive position in AI answers is set by what assistants say about everyone, not only about you. The sampling protocol, and the finding that is not actionable.
Most alternatives pages rank every alternative below the publisher. Sorting by exit reason instead, and why the page works better when some entries win.
Every factual claim about a competitor has a shelf life. The log that records what you checked and when, and the stale claim that gets quoted back at you.
A comparison page that never loses is not read as confident, it is read as marketing. The anti-fit section, where it goes, and why it is the part that gets quoted.
Nine checks to run against a page you already published, ordered by how cheap the fix is. Most pages fail the same three.
A page with no visible date is a page whose facts have no expiry. Where the date belongs, what a real update means, and the practice that makes dates worthless.
Tables are the most extractable thing on a page and the easiest to build wrong. Column rules, the merged-cell trap, and the footnote that vanishes.
Vague sentences are not cited and overclaims get you quoted being wrong. The four properties of a quotable claim, and the hedging that removes all of them.
Answer-first writing is an editorial habit, not an engine requirement. The structure that makes answers easy to find and quote accurately, and what Google says you can ignore.
Ranking well and being cited by an AI assistant are different outcomes produced by different mechanisms. Where the two correlate, where they do not, and what to measure instead.
Tracking whether new pages get indexed one URL at a time tells you nothing. Grouping them into cohorts by publish week shows which batches stalled and why.
The obvious opportunity list is mostly a trap. How to separate pages with a fixable presentation problem from pages ranking for queries they should not.
A domain property mixes your marketing site with preview, staging and app subdomains. How to build a baseline that measures the site you are actually working on.
Search Console's query and page reports do not join the way people assume. What the page-filtered query list actually means, and the three joins that are safe.
Seven questions to answer before telling anyone visibility improved, and the two that most reported lifts cannot survive.
How many times to check a prompt, on what schedule, and how to record it so two periods can actually be compared. The protocol, and the mid-experiment change that invalidates it.
Visibility moved. Was it your content or their model? The control-group method, why you compare the size of the changes rather than their direction, and the assumptions it rests on.
Ask an assistant the same question twice and the sources can differ. The five causes, how to tell which one you are seeing, and what it means for reporting a visibility number.
Any tool that edits your site needs an answer to eleven questions about approval, scope and reversal. The list, why each one matters, and the answers that should end the conversation.
Most AEO trials measure the vendor's demo rather than your market. The fixed-input protocol, what to record daily, and the trap of judging a stochastic system on one sample.
Two product categories, one budget line, and a decision that hinges on where your work actually stalls. The four questions that sort it, with the honest answer when both are wrong.
Social listening, SERP monitoring and AI-answer tracking are three different products often shopped as one. What each actually watches, and which one answers 'are we in the answer?'
Peec, Profound, Scrunch, a manual baseline and Fenn as Otterly alternatives — matched to the three things teams outgrow: prompt volume, reporting depth, or the gap between knowing and shipping.
Before any autonomous SEO tooling touches a client's site: the access, baseline, boundary and rollback checklist — with the missing-baseline failure.
The two-ledger reporting posture that survives bad quarters: a work ledger you control, an outcomes ledger you observe, and dates as the only bridge between them.
The one-page weekly report that makes AI visibility legible to clients: five sections, named fields, a filled example, and the screenshot-theater failure.
Three clients, three risk tolerances, one agency shipping changes. The approval workflow that keeps every change client-attributable, with the batch-approval failure.
AgencyAnalytics, Looker Studio, DashThis, SE Ranking and Fenn compared for agency reporting — with the AI-visibility gap most reporting stacks haven't noticed yet.
Your Search Console data already lists the questions buyers ask that you half-answer. The five-column brief that turns impressions-without-clicks into AEO content work.
When a buyer question is half-answered by an existing page, updating usually beats publishing. The decision table, the URL-equity math, and the cannibalization trap.
GEO, AEO, LLM SEO — same discipline, three names. The tools that actually do the work in 2026, sorted by the four jobs, with the free baseline included.
Profound, Otterly, Scrunch, manual tracking and Fenn as Peec alternatives — who genuinely needs more than Peec, who needs less, and who needs a different kind of tool entirely.
Peec, Otterly, Scrunch, manual tracking and Fenn as Profound alternatives — organized by the actual reason teams shop: price, complexity, or wanting action instead of analytics.
One file governs whether search and AI crawlers can read you — and it's usually years stale. The review checklist: agent inventory, path audit, the noindex trap, and a worked review.
llms.txt is a reasonable courtesy file and a terrible success metric. What it actually does, what no file can do, and the checklist that measures the thing people think llms.txt measures.
Before optimizing for AI answers, verify retrieval: user-agent fetches, CDN and WAF interference, JS-rendering gaps, and a worked check on one URL with the silent-block failure.
Answer engine optimization needs four tool jobs: visibility measurement, technical readiness, content operations, and evidence logging. The honest stack for each job, free options included.
Peec, Profound, Otterly, Scrunch, manual tracking and Fenn compared for AI search visibility — what each actually measures, pricing shapes, and how to evaluate any of them fairly.
When automation touches your site, 'what changed, who approved it, and what happened next' must be answerable in minutes. The audit trail template, a filled example row, and the git-is-enough failure.
A tiering of SEO changes by blast radius and reversibility — what's safe to automate with sampling, what needs per-change review, and what should never ship without a named human owner.
'We can revert' is not a rollback plan. What a tested rollback for metadata changes includes — state capture, the revert path, cache and recrawl consequences — on a worked example.
SEO agent PRs need a different review than human PRs — pattern-level reading, sampled verification, and semantic checks CI can't do. The method, on a worked example.
The checklist a human runs before any AI-proposed SEO change ships — scope, diff, blast radius, rollback — walked through on a worked title-and-canonical example.
A claim that an engine cited you should survive an audit. The log template — fields, storage rules, a filled fictional row — and the paraphrase failure that makes evidence worthless.
Share of voice is only meaningful when the denominator is declared: prompts × engines × runs, stated with the number. The formula, a worked fictional example, and the inflation failure.
How to build the fixed prompt set your AI-visibility measurement depends on — intent mix, inclusion rules, versioning, and the brand-leakage failure that invalidates results.
Mentions and citations are different events with different causes and different value. How to log them separately, with a filled example and the conflation failure to avoid.
A worksheet for measuring where a new SaaS site actually stands in AI answers before you change anything — fixed prompt set, two sampling windows, declared fields.
Visibility Score, Share of Voice, Average Position and Sentiment: what each one measures, how we compute it, and the failure mode of reading it alone.
AEO is the practice of getting your brand cited inside AI-generated answers. Here is what it measures, how it differs from SEO, and what actually moves it.
Every plan tracks your prompts across all five engines, with the MCP server enabled so you can ask about it from Claude or ChatGPT. See what each plan includes.