Starbucks lost 23 points in AI answers.
It happened in one week. We have the receipts.
Between January and July 2026, AI Rankia scanned the AI engines on two brand-free market questions — best coffee shops for remote workers, best coffee near Central Park — and stored every answer. In early June the curve fell off a cliff. Below: the week it happened, the engine where it happened, and the source whose collapse explains it.
The slide, month by month.
Share of voice: how often the brand appears in AI answers to two market questions it does not own. Monthly averages, 1,982 stored runs.
Jan 58% · Feb 64% · Mar 59% · Apr 49% · May 57% · Jun 33% · Jul 37%. April dipped and recovered. June didn’t.
The cliff, scan by scan.
Daily share of voice, May 23 to June 16. The last week of May sits at 70%. The first June scans never get back there.
No scans were stored June 1–9; the drop registered at the first June scan and held. Every point is a stored run you can audit.
Where the drop happened.
Presence rate per engine, before (Jan–May, n=97) vs after (Jun–Jul, n=27). This was not a general fade — it was localized.
| Engine | Before | After | Change |
|---|---|---|---|
| Google AI Mode | 84% | 15% | -69 |
| Claude Search | 17% | 0% | -17 |
| Perplexity | 21% | 4% | -17 |
| Gemini | 47% | 85% | +38 |
| ChatGPT Search | 0% | 0% | +0 |
The cause is in the citations.
Same windows, every citation counted from the stored runs: 15,447 before, 3,433 after. The engines changed their diet.
| Source | Share before | Share after | Change |
|---|---|---|---|
| 12.1% | 1.4% | -10.7 pts | |
| 2.3% | 0.5% | -1.8 pts | |
| 12.3% | 7.1% | -5.2 pts | |
| 5.6% | 11.9% | +6.3 pts | |
| 0.0% | 0.1% | +0.1 pts |
The proof: which pages carried Starbucks — and which never did.
1,847 stored runs across both questions; a text-scan finds the brand’s name in 18% of the answers (stricter than the visibility metric above). For every stored run we check two things in the same answer: which sources it cites, and whether the brand’s name appears in the answer text. Comparing runs where a source is present vs absent isolates each source’s effect. Correlation measured across runs — not a causal claim — sources with enough runs to matter only.
| Source cited in the run | Runs | Brand in answer when cited | When absent | Effect |
|---|---|---|---|---|
| 42 | 100% | 16% | +84 pts | |
| 58 | 91% | 16% | +75 pts | |
| 20 | 95% | 17% | +78 pts | |
| 29 | 93% | 17% | +76 pts | |
| 278 | 0% | 21% | -21 pts | |
| 247 | 0% | 21% | -21 pts | |
| 262 | 1% | 21% | -20 pts | |
| 191 | 1% | 20% | -19 pts |
Borrowed visibility is the whole story.
Of 15,447 citations in the good months, exactly 2 pointed at starbucks.com. The rest was rented ground: Yelp, Reddit, Tripadvisor, city guides.
| Window | Total citations | starbucks.com | Own share |
|---|---|---|---|
| Jan–May (58% visibility) | 15,447 | 2 | 0.01% |
| Jun–Jul (35% visibility) | 3,433 | 3 | 0.09% |
The citation graph also narrowed: 1,065 unique domains cited before, 582 after. New winners in the answer: atly.com, dallasites101.com, sipandco.co, seattle.eater.com — none of which existed in the January mix.
What a weekly alert would have said.
This is the part that is a product, not a story. The same data, run through watch rules:
| Date | Alert that would have fired |
|---|---|
| Jun 10 | Share of voice −40 pts vs 4-week average on “best coffee shops for remote workers” |
| Jun 10 | Google AI Mode presence dropped below 20% (was 84%) |
| Jun 10 | Top citation source yelp.com fell out of the answer set |
The playbook to reverse it.
Each failure above maps to one tool in the suite.
| The problem | The move | In the suite |
|---|---|---|
| 0.01% own citations | Answer-first pages on starbucks.com for the questions the engines actually run | Content Hub |
| One source carried everything | Diversify the citation graph — earn presence in the guides and press the engines quote | Get Cited |
| One engine carried everything | Per-engine tracking and readiness — find why AI Mode dropped you | LLM Readiness |
| Nobody saw it for weeks | Watch rules on share, engines and sources — alert the week it moves | Agents |
More cases, same measurement.
Three brands that went the other way. Click through.
The graph-widening winner: citing domains ×2.6, citations from sources naming the brand +76%.
92% visibility with zero own citations — its firewall blocks GPTBot and PerplexityBot outright.
Even apple.com is 0.2% of market citations. Reviewer sites write the AI’s answer.
Run the same measurement on your brand.
Free scan, live in ~15 seconds. Your visibility, your gaps, the sources AI cites instead of you.
Scan my brand free → Book a demoObservational study of public AI answers, measured weekly by AI Rankia (Jan – Jul 2026). Per-engine presence rates; early-window samples are smaller. Brand names are trademarks of their respective owners, referenced for identification only; no affiliation or endorsement implied.