Automotive OEM GEO Case Study: Recovering AI Citations After a Model-Year Refresh

⚠️ Composite case study — synthesized from public patterns; not a verified single-company case.

TL;DR

A model-year refresh is an entity-drift event for AI search engines. They lose the bridge between the outgoing year's vehicle entity and the new one, so citations that used to surface OEM pages start surfacing third-party listings, dealer aggregators, and review sites. Recovery hinges on keeping URLs, canonical identifiers, and Vehicle schema continuous — not on publishing more content.

Why model-year refreshes break AI citation share

Around 30% of in-market vehicle buyers used an AI tool during research in 2025, and ChatGPT alone captured 68.4% of that AI-tool usage. AI Overviews now appear on roughly 47% of dealer-relevant searches and reduce organic click-through rates by 34-65%. Adobe research with IBM and Profound found that bmw.co.uk went from triggering AI Overviews on 1,284 keywords in January 2025 to 7,134 in October 2025 — a 450% increase in under a year — and projects up to one in five auto searches will be AI-powered by the end of 2027.

The model-year refresh is the single largest entity-disruption event on an OEM site:

• Vehicle Detail Pages (VDPs) move to new URLs that include the new model year.
• Trim names, MSRPs, EPA estimates, and standard equipment change.
• Press releases, brochures, and dealer feeds publish before AI crawlers reindex.
• Outgoing inventory pages are 301-redirected en masse, often to a new lineup landing page rather than a 1:1 successor.

AI engines treat the OEM's "2026 Sedan" and "2027 Sedan" as related but distinct entities. When the outgoing entity's URLs disappear and the incoming entity has a thin authority footprint, citation share collapses for queries that don't pin a year ("best midsize sedan," "Sedan towing capacity," "is the Sedan reliable").

The composite case: what we observed

This case study reconstructs a representative pattern from a mid-volume non-luxury manufacturer (~250k US units/year) across a single 2026→2027 model-year refresh. Numbers are directional and based on publicly disclosed citation tracking, dealer-network analytics, and industry benchmarks; no internal OEM data is reproduced.

Pre-refresh baseline (Q4 2025):

• 38% share of citations across 200 tracked model-level queries on ChatGPT and Perplexity.
• Vehicle schema present on 84% of VDPs.
• 92% of model URLs stable for 18+ months.

Post-refresh trough (week 4 after launch):

• Citation share dropped to 11% across the same 200 queries.
• 41% of AI answers cited third-party listings (Cars.com, Edmunds, KBB) where the OEM had previously been cited.
• 27% of AI answers carried outdated MSRP or trim names.
• Dealer aggregator pages (which had not changed URLs) were now the dominant citation source.

Recovery target: Restore citation share to baseline within one quarter without restructuring the marketing site or running a high-volume content campaign.

What broke: three diagnosable failure modes

1. URL churn severed entity continuity

The OEM's URL pattern embedded the model year. The launch shipped with bulk 301 redirects from /sedan-2026/* to /sedan/, collapsing eight trim sub-pages into a single configurator. AI engines that had cached the outgoing URLs now resolved most queries against archived snapshots or third-party mirrors.

2. Vehicle schema discontinuity

The new VDPs shipped with valid Car and Vehicle schema, but vehicleModelDate, productionDate, and sku properties changed without preserving any cross-year identifier. Schema.org's Vehicle type supports vehicleIdentificationNumber, modelDate, and releaseDate, and Google's vehicle-listing structured data program documents how dealerships maintain inventory continuity across feeds. Without a stable identifier — typically a manufacturer-defined model code or canonical concept ID — AI crawlers had no way to merge entities across years.

3. Authority signals fragmented

Press coverage referenced "the all-new 2027 Sedan" while the OEM's own pages referenced "the Sedan." Knowledge-graph analyses of automotive entity complexity have shown that an Audi A6 (C8) alone has more than 10^33 possible configurations, so AI systems lean heavily on canonical identifiers and clean entity definitions to keep trims, packages, and years aligned. OpenAI's own work on type-based disambiguation describes how models play "20 questions" against discovered categories when text is ambiguous, which makes consistent entity tagging across owned media materially important.

The 90-day recovery playbook

Phase 1 — Stabilise the URL graph (Week 1-2)

• Replace bulk redirects with 1:1 mappings. Each outgoing trim VDP gets a permanent 301 to its closest successor. The configurator becomes a fallback only for retired trims.
• Keep year-stamped URLs alive as canonical archives. Publish /sedan-2026/ as a frozen reference page with a one-line "Replaced by 2027 model" notice and to the successor.
• Stop reusing trim slugs across model years. Adopt a year-agnostic primary URL (/sedan/) plus year-specific archives (/sedan/2027/).

Phase 2 — Restore schema continuity (Week 2-4)

• Add a manufacturerModelCode identifier (or equivalent productID) that persists across model years. This is the single most important field for AI entity reconciliation.
• Populate vehicleModelDate, releaseDate, and modelDate consistently on every trim page. Use ISO dates, not marketing labels.
• Mirror dealer-feed VINs into the OEM's own structured data via Vehicle and Offer so AI engines see the OEM as the authoritative source, not just a downstream feed consumer.
• Validate with the Schema.org validator and Google's Rich Results test before any new VDP goes live.

Phase 3 — Reconcile the entity in owned media (Week 3-6)

• Publish a canonical "What changed in the 2027 Sedan" page with side-by-side trim and spec deltas. Answer-first, table-driven, and explicitly named ("2026 vs 2027 Sedan").
• Bridge press kits and the OEM site. Press releases now link to /sedan/2027/ with the same anchor text used on dealer pages.
• Update Wikipedia and Wikidata stubs. Wikipedia is one of the most cited training-data sources for AI engines; confirm the model-year section and Wikidata identifier reflect the new release date and trim list.
• Refresh comparison pages on owned hubs (/sedan/vs/competitor-x/) so the new trim names and prices appear inside answer-first content the OEM controls.

Phase 4 — Recover citation share through monitoring (Week 4-10)

• Track 200+ model-level prompts weekly across ChatGPT, Perplexity, Gemini, and Google AI Overviews. Capture cited domains, MSRP accuracy, trim names mentioned, and whether the answer recommends the OEM.
• Triage by citation gap. Where third-party sites are cited and the OEM is not, the diagnosis is almost always (a) a missing schema field, (b) an absent answer-first paragraph, or (c) outdated press copy.
• Patch the highest-traffic prompt clusters first. "Is the [model] reliable?", "[model] towing capacity," and "[model] vs [competitor]" recover fastest because the OEM can publish definitive, sourced answers.

Phase 5 — Institutionalise the refresh playbook (Week 8-13)

• Add a model-year refresh checklist to the launch runbook: URL persistence, schema continuity, press-kit anchor consistency, Wikidata refresh, comparison-page updates, monitoring window.
• Pre-warm authority signals 30 days before public launch by giving accredited reviewers and trade press the canonical model-code identifier.
• Lock the canonical concept ID in the OEM's content management system so every internal team uses the same string when referring to a trim.

Results pattern at week 13

These figures are directional, but the relative pattern — a sharp four-week trough followed by a roughly nine-week recovery — has been reproduced across multiple OEMs and dealer networks tracking AI citations.

Lessons

1. Treat every model-year refresh as an entity-drift event, not a content launch. The risk is not "do AI engines know about the new car" — it's "do they know it's the same car."
2. Persistent identifiers beat fresh content. A manufacturerModelCode plus stable URL recovers more citations than fifty new blog posts.
3. Press kits are GEO assets. If the press copy and the OEM site disagree on trim names, AI engines will average across both — usually unfavourably.
4. Dealer aggregators are not the enemy. They are a citation surface. Sync VINs and schema with dealer feeds so the OEM is cited alongside them, not displaced by them.
5. Monitor weekly, not quarterly. Citation trough-and-recovery happens in the same quarter as the refresh; quarterly reviews miss it.

Refresh-event checklist

• [ ] Year-agnostic primary URLs in place (/sedan/, not /sedan-2027/)
• [ ] Frozen archive pages for prior model years
• [ ] 1:1 redirect map (no bulk catch-alls)
• [ ] manufacturerModelCode (or equivalent) on every trim
• [ ] vehicleModelDate, releaseDate, and modelDate populated
• [ ] Press kit and OEM site share trim anchor text
• [ ] Wikipedia and Wikidata reflect the new release
• [ ] 200-prompt citation tracker scheduled weekly
• [ ] Comparison pages updated within 14 days of launch
• [ ] Refresh runbook owner assigned in CMS

FAQ

Q: How long does it take an OEM to recover AI citation share after a model-year refresh?

With the playbook above, citation share for tracked model-level queries typically recovers to baseline in 9-13 weeks, with the steepest gains between weeks 4 and 8 once schema continuity and URL stability are restored. Without it, partial recovery often takes more than two quarters and may not reach the prior baseline before the next refresh cycle begins.

Q: Should an OEM keep the model year in the URL?

No. Keep year-agnostic primary URLs (/sedan/) and use year-specific archives (/sedan/2026/) for historical reference. This preserves accumulated authority across model-year refreshes and prevents the bulk redirect events that disrupt AI citation continuity.

Q: Which schema fields matter most for AI citations on vehicle pages?

The continuity-critical fields are a stable identifier (productID or manufacturerModelCode), vehicleModelDate, releaseDate, modelDate, brand, model, and offers. AI engines use the stable identifier to merge model-year variants, and they use the dated fields to attach the right specs to the right answer (Schema.org Vehicle).

Q: Do dealer aggregator citations hurt the OEM?

Not on their own. They become a problem only when they replace the OEM as the cited source. The fix is for the OEM to publish the same answer-first content (trim comparisons, MSRP tables, towing specs) with structured data on its own domain, so AI engines have a more authoritative option to cite.

Q: Is this playbook only for OEMs, or does it apply to dealer groups?

The same five phases apply to large dealer groups, with two adaptations: inventory-page schema must include VIN-level identifiers, and the citation-monitoring panel should include local prompts ("best Honda dealer in Austin"). Dealer-side citation analysis playbooks documented by industry analysts mirror the URL, schema, and monitoring sequence used here.