In June 2026 Google added the report SEOs had asked for since AI Overviews launched: a Search generative AI performance report in Search Console. Open it and the first question is obvious: where are the clicks?
Right now there aren't any. The report shows:
- Impressions in AI Overviews, AI Mode and generative AI features in Discover
- broken down by page, country, device and date
It does not show clicks, CTR or queries. Several practitioners also report (most recently re-checked in August 2026) that it isn't exposed in the Search Analytics API or the BigQuery bulk export. It's UI only. Google has said more metrics will come, with no date.
You can still get useful answers out of it. This post joins the UI export with the regular Performance data you can get from the API, and estimates which pages AI features are most likely changing your clicks on.
The idea in one paragraph
AI feature impressions are also counted inside the normal Web search type. So for each page you know:
- total Web impressions (API)
- AI feature impressions (UI export)
Divide them and you get an AI exposure ratio per page. Then ask whether pages with high exposure have seen their CTR fall more than pages with low exposure over the same period. That isn't proof, but it's the best signal available until Google ships clicks, and it tells you which pages to look at by hand.
Step 1: Export the AI report
In Search Console: Performance → Search generative AI (the label may vary slightly by account) → Pages tab → set the date range → Export → CSV.
Two notes:
- Data reportedly starts around May 18, 2026, with no backfill. Use a range that starts after that.
- Use exactly the same date range in step 2. Mismatched dates are the most common bug in this kind of join.
Save it as ai_pages.csv. The column names depend on your account's language, so the script lets you set them.
Step 2: Pull Web data for the same pages from the API
Create a service account, add its email as a user on your Search Console property, then:
pip install google-api-python-client google-auth pandas # gsc_ai_exposure.py import sys import pandas as pd from google.oauth2 import service_account from googleapiclient.discovery import build KEY_FILE = "service-account.json" SITE = "sc-domain:example.com" # or "https://www.example.com/" AI_CSV = "ai_pages.csv" AI_PAGE_COL, AI_IMPR_COL = "Top pages", "Impressions" # adjust to your export creds = service_account.Credentials.from_service_account_file( KEY_FILE, scopes=["https://www.googleapis.com/auth/webmasters.readonly"]) gsc = build("searchconsole", "v1", credentials=creds) def web_pages(start, end): rows, offset = [], 0 while True: resp = gsc.searchanalytics().query(siteUrl=SITE, body={ "startDate": start, "endDate": end, "type": "web", "dimensions": ["page"], "rowLimit": 25000, "startRow": offset, }).execute() batch = resp.get("rows", []) rows += [{"page": r["keys"][0], "clicks": r["clicks"], "impressions": r["impressions"], "position": r["position"]} for r in batch] if len(batch) < 25000: return pd.DataFrame(rows) offset += 25000 def main(cur_start, cur_end, prev_start, prev_end): ai = pd.read_csv(AI_CSV).rename(columns={AI_PAGE_COL: "page", AI_IMPR_COL: "ai_impressions"}) ai["ai_impressions"] = pd.to_numeric(ai["ai_impressions"].astype(str).str.replace(",", ""), errors="coerce") cur = web_pages(cur_start, cur_end) prev = web_pages(prev_start, prev_end) df = (cur.merge(prev, on="page", how="left", suffixes=("", "_prev")) .merge(ai[["page", "ai_impressions"]], on="page", how="left") .fillna({"ai_impressions": 0})) df = df[(df.impressions >= 500) & (df.impressions_prev >= 500)] # ignore noise df["ai_exposure"] = (df.ai_impressions / df.impressions).clip(upper=1) df["ctr"] = df.clicks / df.impressions df["ctr_prev"] = df.clicks_prev / df.impressions_prev df["ctr_change"] = df.ctr - df.ctr_prev df["pos_change"] = df.position - df.position_prev df["bucket"] = pd.cut(df.ai_exposure, [-0.01, 0.1, 0.4, 0.7, 1.0], labels=["<10%", "10-40%", "40-70%", ">70%"]) # Rough "missing clicks": what the page would get at last period's CTR df["est_clicks_lost"] = ((df.ctr_prev - df.ctr) * df.impressions).clip(lower=0).round() print("CTR change by AI exposure (median, pages with stable position):\n") stable = df[df.pos_change.abs() <= 1] print(stable.groupby("bucket", observed=True).agg( pages=("page", "count"), ctr_prev=("ctr_prev", "median"), ctr_now=("ctr", "median"), ctr_change=("ctr_change", "median"), ).round(4).to_string()) print("\nPages to look at first (high exposure, stable position, CTR down):\n") print(stable.query("ai_exposure >= 0.4 and ctr_change < 0") .sort_values("est_clicks_lost", ascending=False) [["page", "ai_exposure", "ctr_prev", "ctr", "position", "est_clicks_lost"]] .head(20).round(3).to_string(index=False)) df.to_csv("ai_exposure_by_page.csv", index=False) if __name__ == "__main__": main(*sys.argv[1:5]) Compare the current period with the same dates a year earlier, which removes most seasonality:
python gsc_ai_exposure.py 2026-09-01 2026-09-30 2025-09-01 2025-09-30 Step 3: Read it carefully
Example output (illustrative):
CTR change by AI exposure (median, pages with stable position): bucket pages ctr_prev ctr_now ctr_change <10% 212 0.0410 0.0398 -0.0010 10-40% 96 0.0385 0.0331 -0.0047 40-70% 41 0.0362 0.0254 -0.0101 >70% 18 0.0340 0.0191 -0.0139 If you see a staircase like that, where CTR drops more as AI exposure rises with position held roughly constant, AI features are the likely explanation for the CTR loss on those pages.
Why the filters matter:
-
pos_change <= 1. If a page fell from position 3 to 8, its CTR dropped for ordinary reasons. Removing those pages isolates the AI effect. - 500+ impressions in both periods. Small pages have noisy CTRs that swamp the signal.
- Medians, not means. One viral page shouldn't decide the result.
What it can't tell you:
- Whether a user saw your link inside the AI answer and clicked it. Those clicks are already in your normal click totals and can't be separated out.
- Anything at the query level. The AI export has no queries, so you can't tell which searches triggered the AI feature.
- Causation. Other things change in a year: SERP layout, competitors, brand demand. Treat the output as "pages to look at", not as a number for your CFO.
Step 4: What to do with the list
For each page in the "look at first" list:
- Search its top queries yourself (from the normal Performance report) and see what the AI answer says. Is it accurate? Does it link you?
- If the answer fully satisfies the query, that click may be gone for good. Decide whether the page should target a deeper intent (comparison, implementation, pricing) that a summary can't satisfy.
- If the AI answer links a competitor for a fact you state, check whether your page states it as plainly. Usually it doesn't: the fact is buried in paragraph six or rendered by JavaScript.
- Re-run monthly. Watch the exposure ratio as well as CTR. A rising ratio on a page means Google is showing AI features for more of its queries.
If you'd rather track AI visibility across Google, ChatGPT, Perplexity and Claude in one place than in a CSV, Vista AI's brand visibility dashboard does that. For Google alone, this script plus a monthly calendar reminder covers most of it.
Quick checklist
- [ ] AI report exported (Pages tab) for a clean date range after mid-May 2026
- [ ] Same date range pulled from the API (type
web, dimensionpage) - [ ] Compared with the same period last year
- [ ] Pages with big position changes filtered out
- [ ] Top 20 "high exposure, CTR down" pages checked by hand
Next in the series: Does Claude cite dev.to? A reproducible experiment you can run on your own posts.