QWE AI Academy Your brand can vanish from AI answers overnight – and most analytics never notice
People ask ChatGPT, Perplexity, Gemini, and Google AI Overviews for picks the way they used to type into Google. Stop getting named or cited and pipeline thins out while classic SEO charts stay green. Tracking AI search is how you notice before the quarter is already gone.
Rank trackers do not save you here. Answers are generated, not slotted into a stable top-10. Same prompt, different day, different brands. Free ChatGPT sessions often send no referrer. Google only recently broke out generative AI impressions – and still without clicks.
Below is a hybrid starter you can run this week, mostly free: GSC baseline, a frozen multi-sample prompt sheet, a GA4 channel that catches what little referrer survives, optional API glue, then the noise and fee traps, and a clear stop line so you do not buy a dashboard you do not need.
Hands-on: your first AI search tracking setup
I tried pure manual, light automation, and jumping to a paid seat. Beginners who want to feel the signal before a card charge should stay hybrid. This is the order that wasted the least time for me.
1. Turn on the free Google signal
In Search Console, open Performance and find the Generative AI report (Search and Discover variants). Google launched those reports 3 June 2026; the global rollout note is 31 August 2026. If the property has enough data, you should see impressions for URLs inside AI Overviews, AI Mode, or Discover AI – by page, country, device, date.
What you will not see: clicks, CTR, queries, or a clean split by surface. Export the last 28 days. That file is your only first-party Google AI baseline. Pair it with everything below or you are reading half a story.
2. Build a fixed prompt panel and sample it properly
Ten to twenty buyer questions. Freeze the wording. Three buckets work: brand (“Is [product] good for X?”), category (“best [category] tools 2026”), problem (“how do I fix [pain] without [competitor]”). Week-over-week only means something if the list does not drift.
Same day, run each prompt three times on ChatGPT (search on and off if you have the option), Perplexity, and Gemini. Sheet columns: date, engine, prompt, mentioned Y/N, citation URL, competitors named, rough position or sentiment.
Why three? One pass is a coin flip. Ahrefs and SEJ-style variance writeups keep landing on the same point – probabilistic generation and batch effects – so mention rates swinging 10-30 points run-to-run is normal, not a crisis. Some separation work on top sites even talks about dozens of answers; you do not need that on day one, but you do need repeats.
Treat the sheet like a poll, not a rank tracker. Four weeks of trend beats any heroic screenshot.
First build: about 45-90 minutes. Steady state: 20-30 minutes a week if you protect the ritual.
3. Catch the clicks you can see in GA4
Most AI product traffic never shows up as a neat referrer. Still worth catching the slice that does.
GA4 Admin → Data display → Channel groups → new group. Channel name: AI Search. Condition: Source matches regex. Starter pattern teams actually ship:
chatgpt|openai|perplexity|claude|anthropic|gemini|copilot|grok|you.com|phindPark that channel above default Referral. GA4 evaluates top-down; order is the whole game. Fresh sessions classify going forward only.
Hard limit, from GA4/AI measurement writeups (including xSeek-style guides): free ChatGPT users often strip the referrer. Plan as if you see roughly 17-20% of interactions that click – not the full conversation graph.
4. Optional light automation when the sheet gets old
n8n (or similar) can read prompts, call OpenAI/Gemini/Perplexity, string-match your brand, write rows back on a cron. Cheap models keep token spend boring.
Perplexity is the fee surprise. Sonar-style request charges sit about $5-$14 per 1,000 requests by context size, on top of tokens – see their pricing docs. Search API is called out around a flat $5/1K. A 25-prompt set × 3 samples × weekly stays low single-digit dollars. Daily, multi-engine, fat context? That becomes a real line item fast.
Rather not glue APIs? Cheapest paid start I still check is Otterly Lite at $29/mo (15 prompts, four core engines, daily) – confirm live on their pricing page; add-ons and limits move. Profound-style starters and SEO-suite AI add-ons often sit near ~$99/mo and climb with engines and prompt caps. Validate the category on the free sheet before that invoice.
Common pitfalls that waste weeks
One answer is not truth. I watched a competitor “vanish” two days, then own the cluster again. Variance. Sample.
Do not mash metrics. A mention inside the paragraph, a GA4 session, and “can the bot crawl us” are three different questions. Rising mentions with flat AI-channel sessions is a common empty celebration.
Frozen prompts or stop. Change one clause and you snapped the trend line. Watch tooling that forces web search on every call too – that can diverge from what a normal user sees when the model would not have searched.
What decent early results look like
Four honest weeks: a stable-ish presence rate per engine on category prompts (many brands land somewhere in a wide 20-60% band once noise settles), a map of which rivals own which clusters, GSC AI impressions up or flat, and AI-channel sessions as the conversion reality check – often thin volume, sharper intent.
One product line I followed sat near zero on comparison prompts, then showed up about two in five runs after a pair of tightly sourced comparison pages engines liked citing. The sheet made before/after boringly obvious. Without the multi-sample baseline they would have high-fived random noise.
Some weeks still feel coin-flippy. That is the medium. Fixed panel + repeats beat a prettier dashboard at the start.
When you should skip AI search tracking
Buyers live in classic Google and never open ChatGPT or Perplexity? Low ROI right now. Zero crawlable content or you block major AI bots? Fix fetchability first – empty logs teach nothing. Tiny local shop with almost no category prompt volume? Local SEO still moves faster.
Skip heavy paid seats until the sheet shows a gap you can act on. Burning ~$99-400/mo on fat prompt packs before you know the category moves is how budgets disappear.
FAQ
How many prompts do I actually need?
Start 10-20 across brand, category, and problem. Past ~30 without automation turns into chores. Under ~8 and you cannot see real shifts.
Does a mention without a click still matter?
Yes. Named at decision time, you become the default in the room. The click is nice; it is not the whole story. I have watched pipeline lean on later brand search even when direct AI referrals stayed modest – people hear the name, then Google it on their own clock.
Can I just use my existing SEO rank tracker?
No. Rank trackers score URL positions in link lists. AI search tracking asks whether your brand (or a page) got selected inside a generated paragraph that may not even link out. Different unit, different variance, different tooling. Suites bolting on AI modules are still doing prompt sampling plus citation parsing under the hood.
Open Search Console. Export generative AI impressions. Write twelve prompts. Triple-sample this afternoon. That baseline is what turns next month’s swings from “maybe” into a decision.