Your rankings look fine. Your impressions are flat or even up. And yet the clicks keep sliding, month after month, with no penalty, no algorithm drama and no obvious culprit. If that pattern sounds familiar, the answer is probably sitting in an answer box you have never measured.
Search engines and assistants now write the answer themselves and link out only when they feel like it. A page can be the single most useful document on a topic, get pulled into an AI Overview, quoted almost word for word, and send you nothing. The ranking report says you won. The analytics say you lost. Both are telling the truth about different things.
This guide closes that gap. It walks through six checks, from a ten minute manual test you can run right now to first party reporting in Search Console and Bing, analytics configuration, and server log forensics. By the end you will know which pages AI engines cite, which prompts trigger those citations, which competitors are taking your place, and what to fix when the answer is none of the above.
Figure 1. Pew Research Center tracked real browsing behaviour across 68,879 Google searches. When an AI summary appeared, clicks to traditional results nearly halved, and almost nobody clicked the sources inside the summary itself.
That last bar is the uncomfortable one. According to the Pew Research Center study of Google users and AI summaries, only 1% of visits involved a click on a source cited inside the AI summary. Being cited is not the same as being visited, which is exactly why you have to measure citations separately from traffic.
Figure 2. Zero click behaviour was already growing before generative answers arrived. AI Overviews accelerated a trend rather than starting one.
First, Get Precise About What You Are Measuring
Most people say “am I showing up in AI” and mean three different things at once. Separating them saves a lot of wasted effort, because each one breaks for a different reason and each one is fixed differently.
| Signal | What it means | What it tells you when it fails |
| Retrieval | An AI crawler can reach your page, render it and read the content. | A technical block. Robots.txt, a CDN bot rule, a firewall, or content locked behind JavaScript. |
| Mention | The assistant names your brand in the answer text without linking to you. | The model knows you exist from third party sources, but your own pages are not winning the source slot. |
| Citation | Your URL appears as a linked source in the answer or the source panel. | Your content was retrieved and judged useful enough to ground a claim. This is the real win. |
| Referral | A human reads the answer, clicks your link and lands on your site. | Citations exist but the answer is satisfying the query completely. Expect low volume and high intent. |
Table 1. Four distinct signals that get lumped together as AI visibility.
A site can be cited constantly and get almost no referral traffic. Another can be mentioned often and never cited, usually because review platforms, directories and Reddit threads are doing the talking on its behalf. Knowing which situation you are in determines whether you need technical work, content work or digital PR.
| Measure citations, not vibes: Run every check logged out and in a private window. Personalisation and history quietly skew results. Use a fixed prompt set so month to month numbers stay comparable. Record the date. AI answers rotate sources constantly, so a single snapshot proves very little. |
Step 1: Build a Prompt Set Before You Test Anything
Testing random questions produces random answers. A fixed set of 20 to 30 prompts, written the way buyers actually phrase things, turns a guessing game into a baseline you can track. Keep it small enough that one person can run it in under an hour.
Pull candidates from three places: your Search Console queries filtered to seven words or more, the questions your sales and support teams answer repeatedly, and the wording competitors use in their own headings. Then spread the set across intent types so you are not only testing the easy branded questions.
| Prompt type | Example shape | Why include it |
| Branded | What is [your brand] and who is it for? | Checks whether the model describes you accurately. Wrong facts here are urgent. |
| Category | Best | The highest value slot and the hardest to win. |
| Comparison | [Your brand] vs [competitor], which is better for [use case] | Shows whose framing the model has absorbed. |
| Problem led | How do I fix [specific problem your product solves] | Where helpful content earns citations without brand awareness. |
| Pricing and process | How much does [service] cost and what affects the price | Commercial intent that assistants answer constantly. |
| Local or niche | [Service] in [city] with [specific requirement] | Thin competition, often the fastest citations to win. |
Table 2. A balanced prompt set covers the questions you want to win, not only the ones you already win.
Step 2: Test Each Platform By Hand
There is no substitute for reading the actual answers. Automated trackers sample a model’s output; they cannot tell you that the summary describes your pricing incorrectly or that a competitor’s comparison page is being treated as neutral reference material. Budget roughly 45 minutes for a 20 prompt set across the major platforms.
Each platform exposes its sources differently, so know where to look before you start.
| Platform | Where sources appear | What to watch for |
| ChatGPT | Inline links plus a sources list, shown when web search is active. | Citations are inconsistent. If none appear, ask which sources were used for the answer. |
| Perplexity | Numbered citations on every response by default. | The cleanest platform for testing. If you cannot win here, retrieval is probably the issue. |
| Google AI Overviews | Links beside the sentence they support, plus an expandable source panel. | Does not trigger on every query. Note separately when no overview appears at all. |
| Google AI Mode | Inline links. Every follow up counts as a fresh query. | Fan out means your page may be cited for a subtopic you never targeted. |
| Microsoft Copilot | Numbered citations and a full source list. | Pairs with first party citation data in Bing Webmaster Tools. |
| Gemini | Linked source cards, but often answers from internal knowledge. | Mentions without links are common. Log them separately. |
| Claude | Citations attached to each web search result. | Check referrals from claude.ai alongside the manual test. |
Table 3. Where each assistant shows its working, and the quirk to account for.
How to log the results
Keep the spreadsheet dull and consistent. One row per prompt per platform, with these columns:
- Date, platform and the exact prompt text
- Did an AI answer appear at all (yes or no)
- Brand mentioned without a link (yes or no)
- Your domain cited with a link (yes or no), and which URL
- Every competing domain cited, in the order shown
- Whether the description of you was accurate, generous or wrong
- A screenshot link, because answers change and disputes are common
Two numbers fall out of that sheet. Mention rate is the share of prompts where your brand is named. Citation rate is the share where your domain is linked. Track both monthly. The gap between them is one of the most diagnostic figures you will collect, because a wide gap means the market talks about you but your own pages are not the source of record.
Step 3: Read Your Google Search Console Generative AI Report
Google spent two years folding AI Overview and AI Mode data into ordinary web totals, which made isolating it impossible. That changed on 3 June 2026, when a dedicated generative AI performance report arrived in Search Console. It reached every site worldwide by 31 August 2026.
Where to find it
- Open Search Console and select the property you want to review.
- Go to Performance, then look for the Generative AI entry in the left menu. Discover has its own separate view beneath it.
- Open the Pages tab. Every URL listed there was shown as a link inside an AI Overview or AI Mode response during the selected period.
- Compare the date range against the same window in the standard Web report.
That Pages tab is the single most useful free artefact in this whole process. It is first party, it covers Google’s full query volume rather than a sample, and it tells you exactly which content Google’s systems treat as groundable.
Know the limits before you build reports on it
| What you get | What you do not get |
| Impressions, broken down by page, country, device and date | Clicks or click through rate for AI surfaces |
| A clean view of which URLs Google surfaces inside AI features | A split between AI Overviews and AI Mode, which are reported together |
| Data that also remains inside your overall Web totals | Query level data showing which prompts triggered the citation |
| A property level opt out control in Settings | API access, so you cannot pull it into BigQuery or a scheduled dashboard yet |
Table 4. The generative AI report is genuinely useful and genuinely incomplete. Plan around both.
One underused number: divide your generative AI impressions by your Web impressions over an identical window to get an AI visibility share. Watch that ratio rather than the raw count, because raw impressions move with seasonality. Google’s own documentation on AI features and your website also confirms something worth repeating to anyone selling you a magic fix: there are no additional technical requirements, no special schema and no AI text file needed to be eligible. A page must simply be indexed and allowed to show a snippet.
| Check one setting while you are in there: Search Console Settings contains a Search generative AI control. If a site is opted out, it receives no impressions or traffic from these features at all. Also confirm you are not blocking snippets. A nosnippet or an aggressive max-snippet directive removes a page from AI eligibility entirely, and it is easy to inherit one from a template or a plugin. |
Step 4: Pull Copilot Citation Data From Bing Webmaster Tools
Microsoft shipped something Google still has not: a report that names the pages an assistant actually cited. It is free, it covers every verified site, and almost nobody in the industry uses it.
The AI Performance dashboard arrived in public preview in February 2026. As Microsoft explained when it launched AI Performance in Bing Webmaster Tools, it reports citations across Copilot, AI generated summaries in Bing and selected partner integrations. Verify your site, open AI Performance, and you get four things.
- Total citations. How often your content appeared as a source in AI answers over the period.
- Average cited pages. The daily average of unique URLs from your site that were referenced.
- Page level detail. Which specific URLs were cited, and how that activity moved day by day.
- Grounding queries. The phrases the AI used when retrieving your content. These are not what users typed, which makes them more interesting, not less.
Grounding queries deserve a close read. They show you the internal shape of the question the model was trying to answer when it reached for your page. If the grounding query is a subtopic you mention in passing, you have found a page worth expanding into something more substantial.
Figure 3. One publisher’s 91 day Copilot data: a single guide captured 69% of all citations, and the top four pages accounted for roughly 90%. Citation share is rarely distributed evenly.
That concentration pattern repeats almost everywhere. A small number of genuinely comprehensive pages carry the entire AI visibility of most sites. The practical implication is that deepening three strong pages usually beats publishing thirty thin ones.
Step 5: Separate AI Referral Traffic in Google Analytics 4
Citations are only half the story. You also want to know what the humans who click through actually do, and by default GA4 scatters those sessions across Referral, Direct and Unassigned. GA4 now has a native AI Assistant channel, but it counts forward only and does not claim every platform, so a custom channel group remains the dependable approach.
Set it up once, in about fifteen minutes
- Go to Admin, then Data display, then Channel groups.
- Copy the default group so you keep all the standard channels, and name the copy something like Default plus AI.
- Add a channel named AI Search. Set the condition to Session source matches regex.
- Paste a pattern covering the platforms that matter, keeping the leading and trailing wildcards.
- Drag the AI Search channel above Referral so AI sessions get claimed first, then save.
Custom channel groups evaluate at query time, which means this reclassifies your historical sessions too. You get a backdated view of AI referral growth from the moment you save it, not just data going forward.
Two caveats worth knowing. A meaningful share of AI referral traffic arrives as Direct, because assistants in apps and sidebars often strip the referrer, so treat your AI channel as a floor rather than a full count. And review the regex every quarter, since new assistants appear faster than any pattern stays current.
Once the channel exists, look past volume. AI referrals tend to be low in count and unusually high in intent, with longer sessions and better conversion rates than average organic traffic. Judge the channel on what it converts, not on how big it looks.
Step 6: Confirm AI Crawlers Can Actually Reach You
If the first five checks come back empty, stop optimising content and look at access. A surprising number of sites are invisible to AI answers for a boring reason: a CDN rule, a security plugin or a well meaning robots.txt edit is returning 403s to the exact bots that build answer indexes.
The crucial distinction is between training crawlers and answer crawlers. Blocking a training bot keeps your content out of future model weights. Blocking an answer bot means that engine cannot cite you today, no matter how good your content is.
| User agent | Operator | Purpose | Blocking it means |
| GPTBot | OpenAI | Model training | Excluded from training data. Live citations unaffected. |
| OAI-SearchBot | OpenAI | ChatGPT search index | ChatGPT cannot surface or cite your pages. |
| ChatGPT-User | OpenAI | User triggered fetch | ChatGPT cannot open a page a user asked about. |
| ClaudeBot | Anthropic | Model training | Excluded from training data. |
| Claude-SearchBot | Anthropic | Claude search index | Claude cannot cite you in search backed answers. |
| PerplexityBot | Perplexity | Search index | Perplexity cannot index or cite your site. |
| Googlebot | Search and AI features | Removes you from Search entirely, AI features included. | |
| Bingbot | Microsoft | Search and Copilot | Removes you from Bing and Copilot answers. |
Table 5. The answer crawlers are the ones that decide whether you can be cited at all.
Three checks, in order
- Read your robots.txt properly. Look for explicit disallow rules naming OAI-SearchBot, Claude-SearchBot or PerplexityBot. Also check whether a managed robots.txt feature at your CDN is adding rules you never wrote.
- Grep your access logs. Filter for the user agent tokens and read the status codes alongside them.
- Verify the hits are real. A user agent string is a claim, not an identity, and a meaningful share of self declared AI traffic is spoofed. Operators publish IP ranges at openai.com/searchbot.json, openai.com/gptbot.json and perplexity.ai/perplexitybot.json for exactly this purpose.
Rows of 403 or 429 responses to an answer crawler are the whole diagnosis. So is a 200 response to a page whose main content only exists after JavaScript runs. Disable scripting in your browser and load a key page: if the substance disappears, crawlers are seeing the same empty shell.
Choosing a Tracking Tool, or Not Bothering
Paid AI visibility platforms run your prompt set across multiple models on a schedule and chart the results. They are a convenience, not a different kind of data, and none of them can see every answer, because answers vary by user, session and moment. Start free and upgrade only when manual testing becomes the bottleneck.
| Source | Cost | What it is genuinely good for |
| Manual prompt testing | Free | Accuracy, tone and competitive context. Nothing replaces reading the answers. |
| Search Console generative AI report | Free | First party proof of which pages Google surfaces in AI features. |
| Bing Webmaster AI Performance | Free | The only free report naming cited pages and grounding queries. |
| GA4 custom channel | Free | Behaviour and conversion of visitors who do click through. |
| Server and CDN logs | Free | Whether answer crawlers can reach you, and whether they are real. |
| Dedicated AI visibility platforms | Roughly $29 to $500 a month | Scheduled multi model tracking, share of voice and competitor citation monitoring at scale. |
Table 6. The free stack answers most questions. Paid tools mainly buy you scale and scheduling.
What To Do When You Are Not Showing Up
Work through the causes in order of how cheap they are to rule out. The expensive fixes are at the bottom for a reason.
| Likely cause | How to confirm it | The fix |
| Opted out of AI features | Search Console Settings, Search generative AI control | Switch the control back to include. |
| Snippet suppressed | Inspect the page for nosnippet, data-nosnippet or a short max-snippet | Remove the directive and request a recrawl. |
| Answer crawlers blocked | robots.txt, CDN bot rules, 403s in the logs | Allow OAI-SearchBot, Claude-SearchBot, PerplexityBot and Bingbot. |
| Content invisible without JavaScript | Load the page with scripting disabled | Server render the main content. |
| Page not indexed | URL Inspection in Search Console | Resolve the indexing issue first. Eligibility starts there. |
| Buried answers | The page never states the answer plainly near the top | Open each section with a direct, extractable answer before the detail. |
| No third party corroboration | Competitors are cited through reviews, forums and roundups | Earn mentions on the sources the models already trust. |
Table 7. Diagnose from cheapest to most expensive. Most empty results are caused by the top four rows.
One more thing worth doing before you change anything: read the competitors who are being cited instead of you. Not to copy them, but to see what shape of content the model reached for. Very often it is a page that answers the question in the first two sentences and then earns its length, rather than a page that warms up for four paragraphs.
A Monthly Routine You Will Actually Keep
Monthly is the right cadence for most sites. Weekly produces noise, because AI answers rotate sources naturally. Quarterly means a drop goes unnoticed for too long.
| Cadence | Task | What you are looking for |
| Monthly | Run the fixed prompt set across your core platforms | Movement in mention rate, citation rate and which competitors appear |
| Monthly | Export the Search Console generative AI Pages tab | New pages gaining AI impressions, and pages that dropped off |
| Monthly | Check Bing AI Performance citations and grounding queries | Subtopics worth expanding into dedicated pages |
| Monthly | Review the GA4 AI Search channel | Conversion rate and session quality, not raw volume |
| Quarterly | Audit robots.txt, CDN rules and crawler status codes | New blocks introduced by plugin or infrastructure changes |
| Quarterly | Refresh the prompt set and the GA4 regex | New assistants, new competitors, shifted buyer language |
Table 8. Six recurring tasks. The whole routine fits comfortably in a half day each month.
| The scorecard that matters Citation rate: share of your prompt set where your domain is linked. Mention to citation gap: how often you are discussed without being the source. AI visibility share: generative AI impressions divided by Web impressions in Search Console. Citation concentration: what share of citations your single best page carries. AI referral conversion rate: what the clicks are worth when they do arrive. |
Frequently Asked Questions
Can I see AI Overview clicks in Search Console?
No. The generative AI report shows impressions only. AI Overview and AI Mode clicks remain blended into your overall Web totals with no way to isolate them, which is the single biggest reporting gap in the current setup.
Do I need llms.txt or special schema to be cited?
No. Google states plainly that there are no additional requirements beyond standard eligibility, and server log studies consistently find AI crawlers rarely request llms.txt. Publishing one does no harm, but treat it as optional housekeeping rather than a lever.
Why does my brand get mentioned but never linked?
The model knows you from third party sources such as review sites, directories and community threads, but your own pages are not winning the source slot. That is usually a content extractability problem on your side and a digital PR opportunity off it.
How often do AI citations actually change?
Constantly. The same prompt can return different sources hours apart, which is why a fixed prompt set tracked monthly is far more meaningful than any single check.
Should I block AI crawlers to protect my content?
Separate the two cases. Blocking training crawlers like GPTBot and ClaudeBot keeps your content out of future models and costs you nothing in visibility. Blocking answer crawlers like OAI-SearchBot and PerplexityBot removes you from those answers entirely. Most publishers want the first and definitely not the second.
Is being cited worth anything if nobody clicks?
Often yes, though you have to measure it honestly. The visitors who do click tend to arrive further along in their decision and convert at noticeably higher rates. Being described accurately to a buyer who never visits also has value, it is simply harder to attribute.
Where To Start Tomorrow
You do not need a platform, a budget or a new job title to answer this question. You need an hour and a spreadsheet.
- Open Search Console, find the generative AI report and export the Pages tab. That is your baseline, and it took four minutes.
- Verify your site in Bing Webmaster Tools if you have not already, then read AI Performance. It is the only free report that names your cited pages.
- Write ten prompts your buyers would actually type, and run them through Perplexity and ChatGPT with web search on. Log what you see.
- Build the GA4 custom channel so next month’s data is waiting for you.
- Grep your logs for OAI-SearchBot and PerplexityBot, and look hard at the status codes.
Do that once and you will know more about your AI search visibility than most teams spending money on it. Do it every month and you will see the shifts early, while they are still cheap to respond to.
The sites winning citations in 2026 are not the ones with the cleverest tactics. They are the ones that answer a real question clearly, let crawlers read the answer, and bother to check whether anyone noticed.
Get real time update about this category directly on your device, subscribe now.