SEO Measurement: Metrics, KPIs and AI Search Data in 2026

How Do You Measure SEO? Business Outcomes First, Then the Search Data Behind Them

Measuring SEO means tracking whether organic search, including AI answers, brings in business results, and reading search data to explain why those results moved.

Every SEO number sits on one of three levels:

• Visibility: how often a site appears, counted as impressions in search results and citations in AI answers.
• Visits: the clicks and sessions those appearances turn into.
• Results: the conversions, sales or leads those visits produce.

Google Search Console and Bing Webmaster Tools report what happens up to the click. The site's own analytics, such as Google Analytics 4 (GA4), reports what happens after it.

To measure SEO, decide first which business results organic search should bring in, such as sales or leads, and track those. Then work back through the visits and the search visibility behind them, because that is where you find out why a result went up or down.

Each number also has a source that sets how far to trust it. What an engine or the site's own analytics reports says what happened; what a tool checks from outside says roughly where a site stands; only a controlled comparison says what a change caused.

A site "SEO score" from a checker tool is something else. It grades pages against the tool's own checklist, which makes it an audit, not a measure of results.

AI search added numbers that didn't exist before 2026, such as AI Overview impressions in Search Console, citations in Bing Webmaster Tools and an AI Assistant channel in GA4. It also made some clicks harder to see.

SEO KPIs vs. SEO Metrics: Outcomes to Track, Diagnostics That Explain Them

SEO KPIs (key performance indicators) are the few outcome numbers that say whether search is paying off, such as conversions and revenue from organic search; every other SEO metric is a diagnostic that explains why those KPIs moved.

The split between outcome KPIs and diagnostic metrics is our framing, not an industry standard. In the table's "Kind of data" column, first-party means reported by an engine or the site's own analytics, and sampled means checked from outside by a tool.

Metric	Level	Role	Where it comes from	Kind of data
Conversions from Organic Search (GA4 calls them key events)	Results	Outcome KPI	GA4	First-party
Revenue or lead value from organic search	Results	Outcome KPI	GA4 or the site's sales system	First-party
Conversions from the AI Assistant channel	Results	Outcome KPI	GA4	First-party
Non-branded clicks	Visits	Outcome KPI	Search Console, branded queries filter	First-party
Clicks, impressions, click-through rate (CTR)	Visibility and visits	Diagnostic	Search Console Performance report	First-party
Sessions from the AI Assistant channel	Visits	Diagnostic	GA4	First-party
Average position	Visibility	Diagnostic	Search Console Performance report	First-party
AI Overview and AI Mode impressions	Visibility	Diagnostic	Search Console generative AI performance report	First-party
Citations, cited pages, Citation Share (a site's share of all citations for a query)	Visibility	Diagnostic	Bing Webmaster Tools AI Performance	First-party
Tracked keyword rankings	Visibility	Diagnostic	Rank trackers	Sampled
Mention rate and citation rate in AI answers	Visibility	Diagnostic	AI visibility tools	Sampled
Indexed pages, Core Web Vitals	Before visibility	Diagnostic	Search Console	First-party
Referring domains	Before visibility	Diagnostic	Link tools	Sampled

Non-branded clicks count as an outcome here because, in Google's description of non-branded queries, they show how new users find a site "without any initial intent to go to your site".

Vendor scores, such as a tool's visibility score, Domain Authority or an "AI visibility score", are that tool's own estimate of something no engine publishes. In our view they belong with the diagnostics at most, never among the KPIs.

Report a few outcomes, such as conversions and revenue from search, and pair each one with the diagnostics that explain it, like clicks, impressions and position. When an outcome moves, the diagnostics show why; on their own, they aren't goals.

Indexed pages and Core Web Vitals work differently: they're a technical floor to keep clear rather than numbers to grow, and the Technical SEO guide sets out how to check them.

First-Party Data vs. Sampled Data: What Search Console, Analytics and Rank Trackers Actually Count

First-party data is what an engine reports about a site's own appearances, or what the site's own analytics records: exact for what it covers, with documented gaps. Sampled data comes from tools that check engines from outside: useful for trends and competitors, never exact.

The dividing test is who collected the number: the engine or the site itself, or a tool from outside. A controlled comparison isn't a third source. It's a way of reading either kind to estimate what a change caused.

Measuring results is a different job from sizing demand. Where search volumes and other keyword data come from is covered in SEO Content Fundamentals.

Numbers an engine or your own analytics report tell you what happened. Numbers a tool collects from outside tell you roughly where you stand. Only a comparison against something that didn't change tells you what a change caused, so know which kind a number is before acting on it.

What Search Console Counts, and What It Leaves Out

Search Console counts clicks, impressions, CTR and average position from Google's own data, and Google documents where those counts fall short:

• Impressions: for most results, an impression counts when a link to the site is on the results page a person loaded, "whether or not the item is scrolled into view", in Search Console Help's definition.
• CTR and average position: CTR is clicks divided by impressions. Average position is based on the site's topmost result each time it appeared.
• Anonymized queries: rare queries, "those that aren't issued by more than a few dozen users over a two-to-three month period", are "always omitted from the tables" but "included in chart totals, unless you filter by query" (Google, 2022).
• Date range: the Performance report opens on the last three months, while Google's guide to debugging traffic drops works from 16 months.
• Branded queries filter: open to all eligible sites since March 2026, it sorts queries with "an internal, AI-assisted system", so "some queries may occasionally be misidentified". It works only on top-level properties, and only for sites with "a sufficient volume of queries and impressions" (Google, 2025).

Our takeaway: a baseline has to be built on purpose, over a wider date range than the default, with totals taken from the chart rather than summed from the table rows.

Search Console Clicks vs. Analytics Sessions: Why the Two Never Match

Search Console and Google Analytics never agree exactly: Google says the tools "use different metrics and systems, which means the data won't match completely", and even the closest pair, Search Console clicks and Analytics sessions (visits as Analytics counts them), "are calculated differently".

Five of the eight reasons Google lists, from its page on using the two together:

• Implementation: pages without the analytics tag record no sessions, while Search Console still counts their clicks.
• Consent: visitors who opt out of analytics cookies may be missing from Analytics, which Google says "can skew Google Analytics data"; Search Console still counts their clicks.
• Time zone: Search Console reports in Pacific Time; Analytics uses the time zone set for the property.
• Attribution: Analytics offers three attribution models, the rules that decide which channel gets credit for a conversion, while Search Console counts every click.
• Bots: Analytics "automatically excludes traffic from known bots and spiders", while Search Console "doesn't necessarily filter them out".

Google's own advice is to compare the two as trends, not totals: "what's important is that the general trends have the same pattern". Our rule of thumb follows from it: if both move the same way, the movement is real; if only one moves, check tracking first.

Rank Trackers vs. Average Position: A Sampled Rank and a Reported Average

A rank tracker reports the position it found when it checked a query from one location and device; Search Console's average position averages a site's topmost position across the searches real people ran. They measure different things, so they rarely agree.

The likely causes are ordinary: results vary by location and person, a tracker checks only the keywords someone chose to track, and Search Console averages only over searches where the site appeared at all.

Each answers its own question. A sampled rank shows competitors' positions and keywords a site doesn't appear for yet. A reported average shows the site's real exposure in Google.

Google's spam policies also cover how ranks are collected. They list "scraping results for rank-checking purposes or other types of automated access to Google Search conducted without express permission" as machine-generated traffic, which they say violates Google's spam policies and Terms of Service.

For position data on Google, then, the first-party alternative is Search Console's average position.

AI Visibility Tools: Sampled Prompts and Directional Numbers

AI visibility tools run sets of prompts through ChatGPT, Google's AI features and other assistants and count how often a brand is mentioned or cited, so their numbers are samples of answers that change from run to run.

The Interactive Advertising Bureau (IAB), an advertising trade body, named the common metrics in guidance released in August 2026: mention rate, citation rate and share of voice, with sentiment under what it calls portrayal.

Share of voice is a site's share of the visibility, mentions or citations across a set of queries or prompts, measured against named competitors. Its value depends on which queries or prompts are in the set.

The answers behind these numbers vary widely. In SparkToro's January 2026 study, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times.

Asked the same prompt again, the tools gave the same list of brands less than once in 100 times, and the same list in the same order less than once in 1,000.

IAB sorts this kind of data into two grades. Directional measurement "Identifies patterns and signals trends" but "is not sufficient for budget allocation or executive strategy decisions".

Decision-grade measurement "Meets a higher standard of rigor across query volume, sample size, prompt type coverage, testing cadence, reproducibility, and platform coverage".

Unless a tool shows how it meets that higher bar, we'd read its numbers as directional: trends across many runs, not a position. A stable "AI ranking" for a prompt doesn't exist to be measured.

No outside tool can see inside Google's ranking or AI systems either, a warning from Google that the AI SEO guide quotes.

Designing prompt sets and tracking mentions and citations over time is the subject of AI Visibility, Mentions & Prompt Monitoring.

What Do Search Console, Bing Webmaster Tools and GA4 Report About AI Search?

As of September 2026, Search Console reports how often a site appears in AI Overviews and AI Mode, but not the clicks or queries behind those appearances. Bing Webmaster Tools reports citations and grounding queries, the searches its AI ran to find the pages it cited.

GA4 separates visits from AI assistants such as ChatGPT, but counts clicks from Google's AI features as Organic Search. No engine reports a site's visibility inside ChatGPT, Perplexity or Claude answers.

Checked September 2026. The last column says whether the data is available through an application programming interface (API), the way software reads it:

Surface	Impressions	Clicks	Queries	Citations	Where it's reported	API
Google AI Overviews and AI Mode	Yes, by page, country, device and date	Only inside Web totals	Not in the AI report	No	Search Console generative AI performance report	No AI-specific type
Generative AI features in Google Discover	Yes, in a separate report	Not covered here	n/a	n/a	Search Console	n/a
Clicks from AI Overviews and AI Mode, after the click	n/a	Counted as Organic Search	n/a	n/a	GA4	As Organic Search
AI assistants such as ChatGPT, Gemini, Claude and Copilot	n/a	AI Assistant channel since May 2026; some land in Direct	No	No	GA4	Yes, as a channel
Microsoft Copilot, Bing AI summaries, partners	No	No	Grounding queries, a subset	Citations, cited pages, Citation Share	Bing Webmaster Tools AI Performance	None announced
Answers inside ChatGPT, Perplexity, Claude, Gemini	No first-party report	Visits: GA4 AI Assistant channel	n/a	Sampled by tools only	Third-party tools	n/a

Read Google's AI Overview and AI Mode impressions as their own trend, and count clicks from those features as Google organic, since Google documents no report that splits them out. Visits from ChatGPT and other assistants get their own analytics channel; visibility inside their answers can only be estimated by outside tools.

Search Console: AI Overview and AI Mode Impressions, Counted as Web Search

Search Console counts appearances in Google's AI features as ordinary Web search data, under rules set out on the same Help page that defines impressions:

• "An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position."
• "Clicking a link to an external page in the AI Overview counts as a click", and Google reports those clicks within the "Web" search type of the Performance report.
• "If a user asks a follow-up question within AI Mode, they are essentially performing a new query."
• The generative AI performance report "includes data from the Web search type in the Performance report", so the same impressions appear in both reports.
• A "multimodal" search type, covering Lens, Circle to Search and image uploads, has been rolling out in both reports since September 24, 2026.

The position rule has a consequence Google doesn't spell out. When a site's links move into an AI Overview near the top of the page, its average position can hold or improve while its clicks fall, so a better average position isn't proof that SEO is working.

The Search Analytics API, Search Console's API for performance data, lists no AI-specific search type. Our inference: through the API, AI Overview and AI Mode data appears only inside the web totals.

Write-ups from early 2026 say Search Console can't show AI Overviews. It has reported their impressions since June 2026, and Google's announcement adds: "As of August 31, 2026, we've rolled out these insights to all websites worldwide."

The same announcement says Google is still working out what data would help most, "such as adding additional metrics over time", which is why the map carries a date.

GA4: The AI Assistant Channel vs. Organic Search

Since May 13, 2026, GA4's default channel group, the set of rules that sorts each visit into a channel, has had an AI Assistant channel, with the medium ai-assistant. Google's release note names "chatbots like ChatGPT, Gemini, and Claude".

The channel definitions name ChatGPT, Gemini, Deepseek, Copilot and Grok as example sources, and draw the line that matters. AI Assistant "excludes Google's AI Overviews and AI Mode"; Organic Search covers non-ad search links "including Google's AI Overviews and AI Mode".

Seer Interactive reports three gaps that Google's pages don't state:

• No referrer: visits that arrive without a referrer, the information a browser passes about where a visitor came from, land in Direct.
• Forward-only: GA4 "won't reclassify your historical sessions", so AI visits from before May 13 stay in Referral or Direct.
• Paid mixed with organic: since ChatGPT introduced ads, Seer treats AI Assistant numbers "as a blend of both organic and paid" until a site has checked its own campaigns.

A tracking parameter is a tag added to a link so analytics can name its source even when the referrer is missing. OpenAI says ChatGPT "automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs". Seer's testing adds a caveat: being a source "does not guarantee UTMs will always be included".

The native channel means a custom channel group, rules a site writes itself to sort AI visits, no longer has to be built from scratch. Seer still keeps one alongside it, to compare with older data and to "catch any platforms or edge cases Google's native channel doesn't yet recognize".

What no channel group can do, in our reading, is pull AI Overview or AI Mode clicks out of Organic Search: GA4 receives them as ordinary Google organic visits.

Bing Webmaster Tools: Citations, Grounding Queries and Citation Share

Bing's AI Performance report, in public preview since February 10, 2026, covers "Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations".

It reports total citations, average cited pages, page-level citation activity and grounding queries, shown for a subset of citation activity rather than all of it. Bing's help page is explicit that a citation "does not represent traffic, clicks, or user engagement".

A June 16, 2026 update, "beginning to roll out in preview", adds intents, topics, a comparison with an earlier period and Citation Share.

Bing defines Citation Share as "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query", and the report "does not expose competitor domains".

Bing frames it as "a more directional view" of visibility over time, "not a ranking system or a competitive scoreboard". It's the only first-party AI share metric we found, and Microsoft hasn't announced an API for AI Performance.

Did the Change Work? Before-and-After Reads vs. Controlled SEO Tests

A before-and-after comparison shows that numbers moved, not why: seasons, Google's ranking updates and shifts in demand move them too. Only a comparison against pages that didn't change, or a forecast of what would have happened without the change, estimates what the change itself did.

The checks before crediting a change, the first three from Google's guide to debugging traffic drops and the rest our additions:

• Seasonality: compare with the same period a year earlier. Google's guide uses 16 months of data and Google Trends for this.
• Ranking updates: check whether a Google ranking update landed in the same window.
• Search types: "Analyze different search types separately", in the guide's words, since web, image and video traffic can move apart.
• Brand: split branded from non-branded queries, so a news story or campaign that lifts brand searches isn't credited to SEO.
• Tracking: check whether the analytics tag or the consent banner changed in the same window.
• The change itself: log it the day it ships, in a change log of what changed and where.

Search Console's custom annotations, added in November 2025, put that log next to the data: a note of up to 120 characters on the performance chart, visible to everyone with access to the property.

When a change needs more than a careful read, there are three ways to measure it:

Method	What gets split	What it measures	How it works	Source
User A/B test (conversion testing)	Visitors, between page versions	What visitors do on the page	No cloaking, rel="canonical" on variant URLs, 302 redirects, "Run the experiment only as long as necessary"	Google Search Central
SEO split test	Similar pages, into control and variant groups	Organic traffic to the changed pages	Changes only the variant group and forecasts its traffic from the control group	SearchPilot, a vendor describing its own method
Forecast-based read of one change	Nothing; uses data the change didn't touch	The change's effect on traffic	Forecasts what traffic would have been without the change, then compares	Brodersen et al., Google, 2015, the method behind CausalImpact

Google's testing rules, such as not cloaking (showing search engines different content from what people see), exist so that user tests don't hurt rankings. They don't describe how to measure an SEO change.

When a test is worth it, in our judgment: for changes rolled out across many similar pages, such as a template or a title pattern, with enough traffic to compare groups. A one-off edit to a single page gets an annotation and the checks before crediting a change, and shouldn't be called a test.

A rise after a change shows movement, not cause. Cause needs a comparison: similar pages left unchanged, or a forecast of what would have happened anyway.

Is SEO Worth It? Measuring SEO ROI, and What the Formula Can't See

Whether SEO is worth it for a site is a calculation: return on investment (ROI) is the value organic search brings in, minus what it costs, divided by what it costs. The value side is undercounted in predictable ways.

The cost side includes people, tools and content. The formula is ordinary arithmetic; Semrush's ROI guide uses it with revenue as the value, so $5,000 of revenue on $2,000 of cost is an ROI of 150%.

What the value side misses:

• AI assistant visits without a referrer, which land in Direct, as Seer Interactive reports.
• Visitors who decline analytics cookies, who may be missing from GA4.
• Appearances with no visit: impressions and citations that inform a buyer who never clicks. This one is our reasoning.
• Credit that goes elsewhere: depending on the attribution model, a sale that began with a search may be credited to another channel. Also our reasoning.

Seer's advice for AI referral numbers fits the whole value side: treat them as "a directional floor rather than a precise count".

Vendors publish multipliers for what AI visitors are worth; Semrush's KPI guide, for one, reports its own research finding the average AI search visitor "4.4 times as valuable" as a traditional organic one, based on conversion rate.

Our advice is to compare conversion rate by channel in the site's own GA4 data instead of borrowing a figure measured on other sites.

An SEO Measurement Plan: Outcome KPIs, a First-Party Baseline, AI Search Data and a Change Log

A measurement plan names the outcome KPIs, records a baseline from first-party data, logs the AI search reports on a fixed schedule, and keeps a change log, so every later movement can be read against something.

Some steps are data work, with repeatable inputs and outputs: pulling reports, splitting them, logging values. Others are judgment calls about the business and about cause.

The split and the step order are our proposed method, not a published one. The order runs from what matters to the business, through the data, to reading changes.

1. Choose Outcome KPIs and the Diagnostics Behind Each

• In: the business's goals and the conversions the site records.
• How: pick a handful of outcome KPIs from the results level and, for each, the diagnostic metrics that explain it.
• Out and done: a KPI list where every KPI has a definition, a source and its diagnostics.
• Type: judgment call.

2. Check That the First-Party Sources Are Connected and Complete

• In: Search Console, Bing Webmaster Tools and GA4.
• How: confirm a top-level Search Console property where the branded filter appears (it also needs enough query and impression volume), a verified site in Bing Webmaster Tools and a GA4 key event for each KPI.
• AI setting: check that the Search generative AI setting, which decides whether the site is eligible for Google's AI features, is set as intended.
• Gaps: note the known ones, such as consent opt-outs.
• Out and done: a source list with access, gaps and the date checked.
• Type: data work.

3. Record a Baseline

• In: 16 months of Search Console Performance data, from the report or through the Search Console API, which an agent can read directly, as the Agentic SEO guide describes, plus GA4's Organic Search and AI Assistant data.
• How: split by search type, branded vs non-branded and page group, and take totals from the chart, since anonymized queries leave the table rows short.
• Out and done: a dated baseline table for each KPI and diagnostic, every value labeled first-party or sampled.
• Type: data work.

4. Record the AI Search Reports on a Fixed Schedule

• In: the generative AI performance report, Bing's AI Performance report and GA4's AI Assistant channel.
• How: record AI impressions by page; citations, cited pages, grounding queries and Citation Share; and AI Assistant sessions and conversions. The generative AI report has no API type of its own, so its numbers come from the report or its export.
• Not here: tracking prompts with sampled tools, which is separate monitoring work.
• Out and done: a dated log, one row per period and source.
• Type: data work.

5. Log Every Change When It Ships

• In: each change to the site: a template, content, a technical fix, a migration.
• How: add a Search Console annotation and a change-log row with the date, the pages affected, what changed and what it's expected to move.
• Out and done: a change log that lines up with the data.
• Type: data work.

6. Read Each Change, and Test the Ones Worth Proving

• In: the change log, the baseline and the latest data.
• How: run the checks before crediting a change, then decide whether the change needs a controlled test. For a test, write the hypothesis, the page groups, the metric and the comparison method.
• Out and done: for each change, a reading (moved or not, likely causes, how sure) or a test plan.
• Type: both. Pulling the comparisons is data work; deciding the cause, and whether to test, is a judgment call.

The plan's main output is one table. Every value in the row below is illustrative:

KPI or metric	Definition	Source	First-party or sampled	Baseline	Where it's logged
Non-branded clicks	Clicks on queries the branded filter marks non-branded	Search Console	First-party	Monthly values for the last 16 months, dated	Baseline table

The fourth column is the rule the plan enforces. Every number carries where it came from, so nobody reads a tool's sample as a count, or a before-and-after move as proof of what a change did.

Turning the plan's numbers into dashboards and reports for people outside SEO is the subject of AI SEO Reporting.

Running this plan on a real site, with its own data, is separate work.

Sources

Sourced claims:

• What are impressions, position, and clicks?, Search Console Help, checked September 30, 2026. Cited for how impressions and average position are counted, and the AI Overview and AI Mode counting rules.
• A deep dive into Search Console performance data filtering and limits, Google Search Central Blog, October 2022. Cited for the definition and handling of anonymized queries.
• Performance report (Search results): Overview and basic setup, Search Console Help, checked September 30, 2026. Cited for the three-month default view.
• Introducing the branded queries filter in Search Console, Google Search Central Blog, November 20, 2025, updated March 11, 2026. Cited for availability, the AI-assisted classification, the top-level property and query volume conditions, and the description of non-branded queries.
• Using Search Console and Google Analytics data for SEO, Google Search Central (updated January 7, 2026). Cited for why the two tools don't match, five of the reasons listed, and the advice to compare trends.
• Spam policies for Google web search, Google Search Central (updated August 28, 2026). Cited for scraping for rank-checking as machine-generated traffic.
• IAB Releases "Measuring Visibility in the AI Era" to Help Brands, Publishers, and Agencies Navigate AI-Powered Discovery, Interactive Advertising Bureau, August 3, 2026. Trade-body guidance, cited for the metric names and directional vs decision-grade.
• NEW Research: AIs are highly inconsistent when recommending brands or products, SparkToro, January 2026. Independent research, cited for the method and how rarely lists repeat.
• Introducing Search Generative AI performance reports in Search Console, Google Search Central Blog, June 3, 2026. Cited for the rollout to all sites and more metrics over time.
• Generative AI performance report (Search), Search Console Help, checked September 30, 2026. Cited for impressions only, the overlap with Web search type data and the export.
• AI features and your website, Google Search Central (updated December 10, 2025). Cited for AI feature traffic reported within the "Web" search type.
• Search Analytics: query, Search Console API reference (updated August 11, 2026). Cited for the search types listed, none AI-specific.
• Announcing web multimodal Search performance reporting in Search Console, Google Search Central Blog, September 24, 2026. Cited for the multimodal search type.
• Default channel group, Analytics Help, checked September 30, 2026. Cited for the AI Assistant and Organic Search definitions and the example sources named.
• What's new in Google Analytics, Analytics Help, release note of May 13, 2026. Cited for the AI Assistant channel's launch.
• Your AI Traffic Is Hiding: A Practical Guide to Tracking ChatGPT, Claude, and More, Seer Interactive, updated June 22, 2026. Vendor, cited for referrer-less visits in Direct, forward-only data, paid and organic blending, the utm_source=chatgpt.com tag, keeping a custom channel group, and "a directional floor".
• Introducing AI Performance in Bing Webmaster Tools Public Preview, Bing Webmaster Blog, February 10, 2026. Cited for the surfaces covered, the metrics, and grounding queries covering a subset of citation activity.
• New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare, Bing Search Blog, June 16, 2026. Cited for Citation Share, how Bing frames it, and the preview rollout.
• Debug Google Search traffic drops, Google Search Central, checked September 30, 2026. Cited for 16 months of data, seasonality, ranking updates and search types.
• Adding context to your Search Console data with custom annotations, Google Search Central Blog, November 17, 2025. Cited for annotations and their limits.
• A/B testing best practices for Search, Google Search Central, checked September 30, 2026. Cited for Google's rules for user tests.
• The Math Behind SearchPilot: How SEO A/B Testing Actually Works, SearchPilot. A vendor describing its own method, cited for control and variant page groups.
• Inferring causal impact using Bayesian structural time-series models, Brodersen et al., Google, Annals of Applied Statistics, 2015. Cited for forecasting what would have happened without a change.
• The ROI of SEO: How to Measure SEO ROI (with Formulas), Semrush. Vendor, cited for the formula and its worked example.
• 18 SEO KPIs to measure organic & AI search performance, Semrush, updated August 31, 2026. Vendor research, cited for the "4.4 times as valuable" figure, based on conversion rate, as an example.
• Publishers and Developers - FAQ, OpenAI Help Center, checked September 30, 2026. Cited for ChatGPT adding utm_source=chatgpt.com to referral URLs.
• AI Performance in Bing Webmaster Tools, Bing Webmaster Tools Help, checked September 30, 2026. Cited for citations not representing clicks or traffic.

Argued, not sourced (this guide's reasoning from the sources above):

• The core claim that first-party data says what happened, sampled data says roughly where a site stands, and only a controlled comparison says what a change caused, as this guide's synthesis.
• The visibility, visits and results ladder, the outcome KPI vs diagnostic split, and treating non-branded clicks as an outcome.
• The first-party vs sampled dividing test, and vendor scores as diagnostics at most.
• Building a baseline over a wider range from chart totals, and checking tracking first when only one of Search Console and Analytics moves.
• Why rank trackers and average position disagree, and what each is good for.
• Reading AI visibility numbers as directional unless a tool shows otherwise.
• That no first-party report separates AI Overview and AI Mode clicks, as a reading of Search Console, GA4 and Google's AI features documentation in September 2026, and the API inference.
• The consequence of the AI Overview position rule for average position.
• That no channel group can separate AI Overview or AI Mode clicks in GA4.
• The branded, tracking and change-log checks, and when a test is worth it.
• The last two ROI blind spots, and comparing conversion rate by channel instead of using vendor multipliers.
• The plan's steps, the data work and judgment call labels, and the step order.