Search Visibility Tracking: How to Measure Brand Presence Across Google AI Overviews and ChatGPT
A practical how-to for building a multi-surface visibility register - with prerequisites, numbered steps, scoring rubrics, and the mistakes that make tracking useless.

What Search Visibility Tracking Means Now That Answers Are Generated
Five years ago, tracking visibility meant watching a keyword move from position 14 to position 6. Today a buyer can ask ChatGPT "which medical aid covers chronic dialysis in Gauteng" and receive a synthesised answer naming two or three schemes - none of which had to rank. Search visibility tracking is the discipline of recording whether your brand appears in those generated answers, whether it is cited or linked, and how that presence correlates with impressions and enquiries.
This guide treats it as an operational routine rather than a dashboard purchase: build a flat visibility register, score it with a three-tier rubric, join it to Search Console, and review it weekly.
The Three Surfaces You Actually Need to Track
| Surface | What you can measure | What it cannot tell you |
|---|---|---|
| Classic blue links | Position, impressions, clicks via Search Console | Whether a buyer ever read the snippet |
| Google AI Overviews | Impressions where an overview appeared (new filter in 2025) | Clicks are not isolated - Google reports overview impressions, not per-overview click-through |
| ChatGPT search | Whether your brand is mentioned, cited, or linked in the answer | No public query-level volume data exists |
AI visibility is not traffic. Reported third-party research found roughly 8% click-through when an AI summary appears versus about 15% without it. Treat a mention as brand exposure first and a click source second.
Why Search Console Alone Cannot Answer the CEO's Question
Search Console answers "how many impressions and clicks did my pages get?" It does not answer "when someone asks ChatGPT to recommend a supplier, do we come up?" A CEO who asks about AI presence and receives a Search Console screenshot is being shown the wrong instrument.
What the AI Overviews Filter Shows - and Does Not
Google added an AI Overviews filter to the Performance report that isolates impressions for queries where an overview was served. It shows whether your page appeared in that context and which queries triggered it. It does not show what the overview text said, whether your brand was named inside the summary, or how many people clicked through from the overview block specifically. For the "are we mentioned?" question you must capture the answer text yourself.
Prerequisites: Five Things That Must Exist Before Tracking Starts
- A verified Search Console property with at least 90 days of history.
- A written list of brand names, trading names, and misspellings you must catch.
- A stable list of product SKUs or service categories (20-50 is enough to start).
- A commercial event you can count - enquiry form, call, order, or test-drive booking.
- One named owner who reviews the register every week without exception.
How to Build a Prompt Register (100-300 Items, Versioned, Frozen for the Quarter)
A prompt register is your fixed test set - the AI equivalent of a rank-tracking keyword list. Aim for 100-300 prompts drawn from real questions: sales-call objections, support tickets, site search logs, and competitor comparison queries. Freeze the wording for the quarter; rewriting a prompt mid-quarter makes every prior score incomparable. Store each row with an ID, the exact prompt text, the surface tested, the date, and the owner.
Defining Your Subject List
The subject list is what you search for inside each answer. For an ecommerce brand selling camping gear, subjects might include the store name plus three hero SKUs. For a lead-gen business, subjects are service categories ("bond origination", "commercial HVAC maintenance"). For a dealership group, subjects are model names plus the group's trading name.
How to Capture a Baseline Week Across Google, AI Overviews, and ChatGPT
Pick one calendar week and record everything before you change a single page. Run every prompt in the register on each surface and save the raw answer text to a dated folder. This baseline is what every later week is compared against; without it, improvement is a feeling rather than a measurement.
Running ChatGPT Prompts Reproducibly
ChatGPT answers vary between runs. Run each prompt three times in fresh conversations, score each run, and record the average. Note the model version and date, because OpenAI ships changes without notice. A single-run read is a coin flip, not a measurement.
prompt_id | prompt_text | surface | runs | scores | avg
P014 | best medical aid chronic dialysis | chatgpt | 3 | 1,0,1 | 0.67
The Scoring Rubric: Mentioned, Cited, and Linked
| Score | Definition | Commercial meaning |
|---|---|---|
| 0 | Absent | No exposure |
| 1 - Mentioned | Brand named in the answer text | Awareness without a path to act |
| 2 - Cited | Named with a source reference | Credibility borrowed from the citation |
| 3 - Linked | Clickable link in the answer | Direct route to site or store locator |
The ordering is deliberate. A mention builds memory, a citation builds trust, and a link builds a session. Scoring all three separately stops you from conflating "we got named" with "we got traffic".
How to Join Visibility Scores to Search Console AI Overview Impressions
Export the Performance report filtered to AI Overviews and pivot it by query. Then match those queries to your register where wording overlaps - exact matches are rare, so match on intent themes ("chronic dialysis cover", "bond pre-approval requirements") and aggregate the impressions behind each theme. The join gives you one number per theme per week that combines off-site presence with on-site demand.
A Worked Example
| Week | AI Overview impressions (theme) | Avg ChatGPT score | Enquiries |
|---|---|---|---|
| Baseline | 4,210 | 0.67 | 11 |
| Week 6 | 6,880 | 1.33 | 19 |
Impressions rising while the score stays flat usually means you are winning classic results, not AI answers. Rising score with flat impressions points to prompts too niche to have measurable demand.


