Your monthly SEO report is probably accurate, and it is also measuring a shrinking share of where buyers now get answers.
It will show backlinks, indexed pages, organic traffic, click-through rate, keyword rankings, domain authority and Search Console impressions. Those seven numbers describe how Google's ten blue links performed. They say nothing about whether ChatGPT, Gemini, Claude or Perplexity named your business when a buyer asked them who to hire.
That gap is what the future of SEO is mostly about. Buyers now ask an assistant a long question and get a short list of names back. If you are on the list, you get the enquiry. If you are not, no report you receive will tell you, because nothing in it was built to look.
Below, you will find the seven numbers you probably see, why each is losing power, and ten measures that sit underneath them, each with a cheap way to check it.
What the seven classic SEO numbers still tell you
The classic seven still describe how well Google can find, rank and send people to your pages. They are worth keeping. They are no longer enough to describe visibility.
Here is each one, and where it is weakening.
- Backlinks. A count of links pointing at your site. It does not show whether anyone who matters is talking about you.
- Indexed pages. How many of your URLs Google has stored. Google's documentation says a page must be indexed and eligible to show with a snippet to appear as a supporting link in its AI features. So indexing is a condition for being cited, not a score.
- Organic traffic. Visits from search. Falling traffic can now mean lost rankings or a search engine that answered the question without sending anyone.
- Click-through rate. The share of impressions that become visits. It drops when an answer appears above your result.
- Keyword rankings. Your position for a tracked list of terms. More on why this misleads below.
- Domain authority. A score from a software vendor, not a number Google publishes. It is a rough proxy for link strength and has no view of AI answers.
- Search Console impressions. Times your page appeared in results. Google says sites shown in AI features are included in overall Search Console traffic under the Web search type, so you cannot separate AI exposure from ordinary exposure in the standard report.
Clicks fall when an AI summary appears
Pew Research tracked 900 US adults in March 2025 and logged 68,879 Google searches. Users clicked a traditional result on 8% of visits that showed an AI summary, against 15% without one. Only 1% clicked a link inside the summary. About 18% of searches produced a summary in that sample.
xychart-beta
title "Share of Google visits with a click (Pew Research, March 2025)"
x-axis ["Traditional result, no AI summary", "Traditional result, AI summary shown", "Link inside the AI summary"]
y-axis "Percent of visits" 0 --> 20
bar [15, 8, 1]
Searchers also left the browser more often after an AI summary: 26% of visits against 16%. Our zero click search guide covers the UK, where SparkToro put the zero-click rate at 69.5% in 2026.
Rankings are a weaker guide to who gets cited
Ahrefs published an update in March 2026, using 863,000 keyword results and 4 million AI Overview URLs. It found 37.9% of cited URLs ranked in Google's top ten for the same query, 31.2% ranked between 11 and 100, and 31.0% ranked beyond 100.
pie showData
title "Where AI Overview citations ranked for the same query (Ahrefs, March 2026)"
"Positions 1 to 10" : 37.9
"Positions 11 to 100" : 31.2
"Beyond position 100" : 31.0
An earlier Ahrefs study put the top-ten share at about 76%. The two are not a clean comparison, because Ahrefs changed its parsing and Google changed the model behind AI Overviews in the same period.
The survey we cover in our guide to what 45 GEO studies found reports a similar audit, with 53% of domains used by AI Overviews outside the organic top ten. A ranking report cannot see any of this.
What changed at the waterline
Three things changed between the numbers you see and the numbers that now drive enquiries.
First, the answer often arrives before the list of links. Pew's click data is the plainest evidence.
Second, the answer is assembled differently. Google describes a "query fan-out" technique, in which AI Overviews and AI Mode issue multiple related searches across subtopics and data sources. A buyer types one question and the system runs several, so you can be cited for a sub-question you never targeted.
Third, the output is a set of named brands. Ranking tells you where you sat on a page of ten. An AI answer gives three or four names and moves on, so the question becomes whether you are one of them.
flowchart TD
A[Buyer types a prompt] --> B[Assistant splits it into sub-queries]
B --> C[Searches run across the web and indexes]
C --> D[Pages and sources are fetched and read]
D --> E[Answer is written from those sources]
E --> F[A few brands are named or cited]
F --> G{Does the buyer click?}
G -->|Sometimes| H[Visit to your site]
G -->|Often not| I[Brand remembered, or forgotten]
Each step is a place where you can be dropped. Our guide to the six stages from prompt to citation walks through them, and the measures below map onto those steps.
The ten measures below the waterline
None of the ten replaces the classic seven. They answer a different question: who gets recommended. One caution applies to all of them. AI answers vary run to run. In one 45-day audit described in the survey, the set of sources cited on a given day overlapped only 34% to 42% with the day before. Any single check is an anecdote, so measure repeatedly, as in our piece on why a single ChatGPT check tells you nothing.
Group 1: presence
Presence measures answer one question: when buyers ask, do assistants say your name?
Share of voice in AI answers. This is the proportion of answers, across a fixed set of prompts, that name you compared with named competitors. Check it by running your prompt set and counting mentions for you and three rivals. What to do: read the answers where you are absent and note what the named businesses have that you lack.
Citation rate across your prompts. This is how often your own URL appears as a source, which differs from being named. An assistant can recommend you without linking, or link you without recommending. Record both. If you are named but not cited, your own page is probably not the one being read.
Visibility by platform. ChatGPT, Claude, Gemini and Perplexity draw on different sources, so you can be visible in one and invisible in another. Check each separately, never as one blended score, and look at the weakest first. Bing indexing matters to ChatGPT, for example, while Gemini leans on Google's own index.
Coverage of your full prompt universe. This is how many of the questions buyers could plausibly ask you appear in your test set, and in how many you show up. Most owners test only the obvious "best X in my town" prompt. Write prompts for each service, audience, comparison and objection. Our 50-buyer-prompt exercise for a trade business shows how to build the list.
Group 2: sources
Source measures show which pages the assistants trust when they answer in your category.
Which URLs get cited in your category. Collect the sources cited across your whole prompt set, and count the domains. You will often find directories, forums and comparison sites ahead of any business website. Treat that list as your target list. Peec, a monitoring vendor, reported that Reddit sat in the top five cited domains on all five engines it tracked in March 2026.
Mentions on editorial and third-party sites. This is how often independent sites discuss you: trade press, directories, review sites, local publications. One study in the survey found ChatGPT recognised 99.4% of the named startups but surfaced them in only 3.32% of open discovery queries. Being known and being recommended are different. See why third-party mentions can matter more than your own site. Earn mentions as yourself, with real expertise, and avoid anyone offering to manufacture them.
Cited and ranking on Google. Compare your cited pages with your Google rankings for the same query. Four outcomes are possible: cited and ranking, cited and not ranking, ranking and not cited, neither. The Ahrefs split above shows why all four occur. For pages that rank but are not cited, make the key passage answer the question directly. For pages cited but not ranking, leave them alone until you know why.
Group 3: perception
Sentiment in AI answers. This is how the assistant describes you when it does name you. Is it accurate, current and favourable? Assistants repeat what the web says, including old complaints, outdated prices and wrong service lists. Ask each platform what your business is like and whether it is any good, then read the sentences closely. Correct anything wrong at its source, whether that is your own site, a directory listing or a review profile. A dental practice with an old fee page, or a law firm listed under the wrong practice area, will see it repeated back.
Group 4: mechanics
Mechanics measures show whether the machinery can reach you at all.
AI crawler visits to your site. Assistants send crawlers to fetch pages. OpenAI documents three: GPTBot for training, OAI-SearchBot for search and ChatGPT-User for fetches a person triggers. Check your server logs, or ask your host, for those user agents over the last 30 days. If search crawlers are absent, check robots.txt and firewall rules. Blocking training is a legitimate choice. Blocking search crawlers removes you from the answers.
Crawling is not the same as sending visitors. Cloudflare Radar data for July 2025, reported by PPC Land, measured how many pages each platform fetched for every visitor it sent back:
| Platform | Pages fetched per referral (July 2025) |
|---|---|
| 5.4 | |
| Perplexity | 194.8 |
| OpenAI | 1,091 |
| Anthropic | 38,065 |
The Anthropic figure is probably an overstatement, because the Claude app sends no referrer header and the visits go uncounted. The point holds anyway: a crawler visit is a sign you can be read, not a sign you will be visited.
Sub-queries behind each prompt. These are the extra searches the assistant runs for a single question, such as "price", "reviews", "alternatives" or "near me". You rarely see them directly. Infer them by reading the sources in each answer and asking what question each one answers. Make sure you have one clear page for each real topic. Google warns that mass-producing a page per fan-out variation can breach its scaled content abuse policy, so do not do that.
Tracked today versus tracked next
This table puts the two halves side by side.
| Tracked today | Tracked next |
|---|---|
| Backlinks | Mentions on editorial and third-party sites |
| Indexed pages | AI crawler visits to your site |
| Organic traffic | Citation rate across your prompts |
| Click-through rate | Share of voice in AI answers |
| Keyword rankings | Cited and ranking on Google, side by side |
| Domain authority | Visibility in ChatGPT, Claude, Gemini and Perplexity |
| Search Console impressions | Coverage of your prompt universe, plus sub-queries and sentiment |
The pairs are loose on purpose. Each right-hand row answers a question the left cannot.
What to do this quarter
You can set this up in a quarter without buying a tool.
- Run a 20-prompt manual test. Write 20 prompts a buyer would type: five on the service, five on comparisons, five on price and risk, five on location or audience. Run each on ChatGPT, Claude, Gemini and Perplexity. Record whether you are named, whether you are cited, who else appears and what the sources are. Repeat each prompt a few times, since answers vary. Our spreadsheet tracker guide gives a layout.
- Set up AI referral sources in GA4. In the traffic acquisition report, filter session source for chatgpt.com, perplexity.ai, gemini.google.com and claude.ai. Expect small numbers, and treat them as a floor, since some assistants send no referrer.
- Check AI crawlers in your server logs. Look for GPTBot, OAI-SearchBot, ChatGPT-User and the equivalents for other platforms. If you cannot read the logs, ask your host or developer for a 30-day extract.
- Ask your agency for four things. The share of voice figure with the number of runs behind it. The cited-URL list for your category. The cited-versus-ranking comparison for your top pages. A sentiment read, with errors flagged for correction.
If an agency cannot show the number of runs behind each figure, you cannot tell change from noise.
Where this leaves your reporting
Keep the classic seven. They tell you whether Google can find and rank you. The ten tell you whether buyers will be told your name.
Nobody has shown that any specific tactic reliably moves these numbers, as our review of the GEO research makes clear. So the first job is measurement: know where you stand before spending on changes.
If you would like to see your own baseline, our AI search visibility service starts with that prompt test, or you can get in touch and we will talk through what to measure first.