Run 3 — Motor, five engines, and an honest number
Run 3 is the first measurement rigorous enough to put in front of a board. We held the question set to Motor only, ran 100 queries across five answer engines, three times each — 1,500 calls, 1,487 clean — counted AXA separately from GIG throughout, and put a 95% confidence interval on every headline. It changed the number, and more importantly it changed the story.
The difference in coverage, models and method
Each run tightened the method. The number fell each time — not because GIG lost ground, but because we removed the things that were flattering it.
| Run 1 | Run 2 | Run 3 (this) | |
|---|---|---|---|
| Scope | Mixed lines | Motor | Motor |
| Queries | ~50, "mentioned anywhere" | 100 (set not retained) | 100, intent-tagged |
| Engines | 3 | 4 (claimed) | 5 |
| Repeats per query | 1 | 1 | 3 |
| AXA handling | counted as GIG | split | split, throughout |
| Confidence interval | none | none | Wilson 95% |
| Headline GIG | ~83% | 65% | 51% |
The five engines, and the models behind them, are not the same kind of thing — and that distinction turns out to be the whole point:
| Engine | Model queried | Reads… | Type |
|---|---|---|---|
| ChatGPT | gpt-4o-mini (no browsing) | its own training memory | parametric |
| Claude | claude-haiku-4-5 (no browsing) | its own training memory | parametric |
| Perplexity | Sonar | the live web, with citations | grounded |
| Google AI Overview | SerpApi AI Overview | Google's live index | grounded |
| Google organic | Serper SERP | Google's live index | baseline |
Share of queries — how the 100 Motor questions split
The set is built from real Motor intent, weighted toward the product and service questions where the money is, not just the easy brand questions.
Queries per intent, of 100 Motor questions. Each runs on 5 engines × 3 repeats.
The performance shift — why 65% became 51%
The drop is rigour, not decline. Three things moved the number down, all of them corrections:
- AXA is no longer counted as GIG. AXA alone appears in 46% of Motor answers — and on Claude, 94%. Earlier runs folded that into GIG. Stripping it out is the single biggest correction.
- The parametric LLMs are now in, and weighted. ChatGPT and Claude name GIG only ~18% of the time from memory; including them honestly pulls the blended average down.
- Three repeats average out the luck. A single pass can catch a brand on a good day; three passes give the steady-state rate and a confidence band instead of a point.
GIG "included" headline as method tightened. The fall is inflation being removed, not visibility being lost.
The split that explains everything
The blended 51% hides the real finding. Split the engines by what they read, and a chasm opens:
| Engine | GIG incl. | AXA | GIG linked |
|---|---|---|---|
| ChatGPT · parametric | 17.7% | 56.3% | 0% |
| Claude · parametric | 18.3% | 94.0% | 0.7% |
| Perplexity · grounded | 62.2% | 16.2% | 65.6% |
| Google AI Overview · grounded | 79.0% | 31.0% | 76.0% |
| Google organic · baseline | 78.3% | 30.3% | 75.0% |
On the engines that read the live web, GIG is named 62–79% of the time and its own pages are cited 66–76%. On the engines that answer from memory, GIG sits at ~18% — and AXA fills the space, peaking at 94% on Claude. Average the two worlds and the gap is stark:
Mean GIG inclusion, parametric vs grounded. The 55-point gap is the whole strategy in one chart.
By intent — where Motor is won and lost
GIG owns the branded and competitive ground (87%, 81%) and holds the category question (56%). It falls away on the high-intent service moments — claims (35%), eligibility and how-tos like black points and traffic fines (29–39%). Those are buyers in-market now, and they are exactly where GIG's content is thin.
GIG's gap is in the models' memory, not on the live web.
On every engine that reads the open web, GIG already wins and its own pages get cited. The weakness is concentrated in the two places that answer from training memory — and there, AXA still owns the space the GIG name should. So the problem is not "GIG is invisible." It is "the rebrand hasn't reached the models' memory yet, while GIG's live presence is already strong."
That splits the work cleanly. Fast lane: feed the grounded surfaces — schema, crawlable prices, the service-query content — where wins land in days (this is the prices-into-Gemini play). Slow lane: rebuild brand memory so GIG, not AXA, is what the models recall — a months-long job measured run over run.