The holiday shopping season’s shiny new helpers are showing an old‑data problem: the leading AI shopping assistants from OpenAI, Google, Microsoft and smaller rivals routinely surface
last‑generation devices unless prompted to fetch “the latest” — a failure that can cost shoppers money, time, and confidence. Real‑world tests across ChatGPT’s Shopping Research, Google’s Gemini agentic features, Microsoft Copilot’s shopping sidebar, and Perplexity’s product search reveal a consistent pattern: assistants present outdated model recommendations with the same confident tone they use for current analysis, leaving the average buyer exposed to stale choices and missed upgrades.
Background / Overview
Conversational shopping assistants are being marketed as a replacement for the tab‑shuffle shoppers tolerate today: ask a question, answer a few follow‑ups, and receive a curated buyer’s guide that compares models, prices and availability. OpenAI introduced
ChatGPT Shopping Research in late November 2025 as a boxed, retrieval‑augmented shopping workflow designed to run follow‑ups, compare SKUs, and return inventory‑aware links. Microsoft folded price history, review summarization, price‑tracking and cashback cues into
Copilot’s shopping sidebar. Google expanded
Gemini into agentic shopping tasks — including a Duplex‑style “let Google call” capability to check local store stock — and Perplexity added integrated shopping workflows with in‑tool checkout options. Independent testing and newsroom reporting confirm these launches and the broad feature sets being promised to users. These products are useful innovations in principle: they reduce friction and can assemble research faster than a casual shopper can by hand. But when the retrieval layer that feeds them lags or omits major sources, the assistant’s
presentation of recommendations becomes misleading — and in practice, several assistants have defaulted to older but well‑reviewed models rather than surfacing the newest releases.
How the assistants performed in testing
The headline failures
- ChatGPT’s Shopping Research often produced thorough buyer’s guides but surfaced older models as top picks in category tests (for example, recommending the Garmin Vivoactive 5 when the Vivoactive 6 was already available).
- Google’s Gemini sometimes favored Pixel Watch 2 (2023) content in results even after Pixel Watch 4’s 2025 release — and its “call for me” / “let Google call” agentic feature produced an embarrassing operational failure in at least one test where an email arrived saying the stores contacted did not stock the model.
- Microsoft Copilot delivered the best interface for shopping work: price history charts, aggregated Amazon review snippets, and price‑tracking alerts in a sidebar were immediately useful — yet Copilot still missed some newer SKUs (for example, surfacing the CMF Watch Pro 2 while missing the CMF Watch Pro 3 in the same exploration).
- Perplexity returned mixed results: fast links and a transactional flow (PayPal integration) are compelling, but the assistant sometimes mixed current picks with much older models (for example, surfacing a 2021 Samsung Galaxy Watch 4 alongside 2025 models).
This pattern is not isolated: the Verge ran a cross‑platform test that reproduced the same trend, concluding that unless the user explicitly asks for
current or
latest models, many assistants fall back to older, well‑rated options.
Real example: smartwatches and the Nothing CMF Phone 1
A practical test highlighted how this fails in a real purchase scenario. When asked to find Android smartwatches compatible with the Nothing CMF Phone 1, the assistants returned a mixture of old and new models:
- ChatGPT centered its recommendation on the Garmin Vivoactive 5 (a strong 2023 model) rather than the Vivoactive 6, which Garmin launched in April 2025 and that reviewers praised for improved GPS, more storage and a new Smart Wake alarm. The Vivoactive 6’s launch and specs are well‑documented on Garmin‑facing sites and mainstream reviewers.
- Google’s outputs tended to surface the Pixel Watch 2 (2023) in some queries even after the Pixel Watch 4 became widely available in 2025; Google’s own Pixel Watch 4 product pages and hands‑on reviews show meaningful battery, charging and processing improvements over the 2023 model.
- Perplexity mixed up‑to‑date picks like the Pixel Watch 4 with much older Galaxy Watch 4 results from 2021 — a poor experience for shoppers who expect a clean shortlist.
- Microsoft Copilot quickly suggested the CMF Watch Pro 2 (a match for the CMF Phone 1) but did not surface the CMF Watch Pro 3 despite that model’s availability and coverage in mainstream outlets.
The upshot: a shopper who relies on a conversational assistant without verifying release dates could easily buy a last‑generation device and miss firmware, battery or compatibility improvements in the newer model.
Why this keeps happening: technical and operational causes
1) Retrieval gaps, permissions and closed marketplaces
Most assistants combine a language model with a retrieval layer (RAG — retrieval‑augmented generation). The RAG layer must either crawl publicly accessible pages or ingest merchant APIs/product feeds. When a major local marketplace blocks crawlers or does not offer a usable feed, the assistant’s coverage narrows and its results skew toward the platforms it
can reach. Independent testing and vendor documentation show this exact weakness in practice (the Korea/Naver example is illustrative).
2) Index freshness and caching windows
Even when crawling is permitted, update cadence matters. Merchant prices, SKUs and product launches change fast. If the retrieval index refreshes weekly or monthly — or a downstream aggregator’s feed lags — the assistant will synthesize recommendations from stale snapshots. Several test reports and product pages warn that price and availability can change faster than the assistant’s index updates.
3) Training data vs. retrieval data confusion
An assistant’s
language model is trained on historical corpora; its
retrieval component is meant to supply current facts. When UIs or outputs don’t distinguish between the two, users see confidently worded guidance that may rely partly on older model priors rather than fresh retrieval—especially for niche or newly released products. Some testers concluded the effective “knowledge cutoff” for product details felt like
months old, though precise training‑cutoff dates are not disclosed by providers. That specific numeric claim (e.g., “18–24 months old”) is plausible in light of model update cycles but not independently verifiable without vendor confirmation; treat such figures as a working hypothesis unless vendors publish cadence details.
Caution advised.
4) Normalization and SKU matching errors
Products are listed under slightly different SKUs, titles and bundle descriptions across retailers. The retrieval layer must normalize those attributes so the assistant can compare like‑for‑like. Imperfect normalization leads to recommendations that conflate models or fail to notice a newer SKU that differs only in a minor title token. This is an engineering challenge in e‑commerce search and affects all players.
5) Legal and scraping constraints
Some merchant sites expressly forbid scraping; when crawlers are blocked, assistants must rely on alternate sources or APIs. That constraint creates uneven coverage across brands and regions unless platforms agree to formal API partnerships or allowlisting. OpenAI and others explicitly document this fallback behavior.
Strengths observed — where assistants already add real value
Despite the recency problem, several genuinely useful capabilities emerged across platforms:
- Conversational disambiguation: follow‑up questions that narrow a shopper’s intent (budget, use case, size) save time and surface better matches than a single keyword search.
- Integrated price signals: Copilot’s price history charts, alerts and aggregated review summaries are practical levers for deal hunters who want one‑stop monitoring.
- Actionable buyer’s guides: ChatGPT’s Shopping Research can create side‑by‑side comparisons and “best if” annotations that are legitimately helpful when the retrieval layer is fresh.
- Agentic convenience when it works: Google’s Duplex‑style calling and agentic checkout can relieve tedious tasks (e.g., calling multiple local stores) if the execution is reliable. Early reporting confirms the existence and promise of these agentic flows.
The consumer risk matrix: what shoppers lose when assistants are stale
- Missed improvements and compatibility pitfalls. Buying last‑generation smartwatches can mean losing better GPS, longer battery life, firmware features or compatibility fixes in the newer model. The Vivoactive 6 and Pixel Watch 4 examples show meaningful incremental improvements that matter to buyers.
- False confidence. Assistants often present recommendations with no visible provenance or confidence score, causing consumers to overweight AI outputs. Several product guides explicitly caution users to verify prices and stock on merchant pages.
- Regional blind spots. In markets dominated by local platforms (e.g., Naver in Korea), omission of a dominant feed can exclude community favorites entirely and skew national recommendations.
- Commercial bias risk. If assistants adopt allowlisting, affiliate links, or merchant partnerships without transparent labeling, merchants that cooperate with platforms could gain disproportionate visibility. That economic incentive could distort otherwise neutral recommendations unless regulated or disclosed.
Recommendations for shoppers (practical checklist)
- Treat assistant output as a shortlist, not a final purchase order. Always confirm release dates and specs on manufacturer pages before buying.
- Ask explicitly for recency: include terms like “latest model,” “released in 2025,” or “newest version” when you want current hardware. Tests show assistants often obey these modifiers.
- Use the assistant to narrow choices (features, budget), then validate on a trusted review site or the vendor’s product page.
- For big purchases, cross‑check at least two independent sources (manufacturer page + major review outlet) before committing. Major claims about launch dates or critical specs should be verifiable in two places.
- Watch for provenance: prefer assistants that surface links and timestamps for the pages they used to generate recommendations. Assistants that show their sources enable faster verification.
Recommendations for vendors, marketplaces and policymakers
- Merchants must publish structured, machine‑readable product feeds and consider formal API partnerships or allowlisting paths; that is the robust way to ensure consistent inclusion in assistant results. OpenAI and others have documented allowlisting procedures and recommend this path for merchant discoverability.
- Platforms should add explicit provenance metadata and confidence signals to assistant outputs: which pages were read, timestamp of retrieval, and a short confidence score would materially improve consumer trust. Several industry analysts and product previews have emphasized this need.
- Regulators and consumer‑protection agencies should require disclosure where recommendations are monetized or where allowlisting privileges ranking. Without disclosure, consumers cannot distinguish ad‑favored placements from organic recommendations.
What vendors should fix first (engineering priorities)
- Shorten retrieval refresh windows where feasible for product catalogs; prioritize real‑time APIs for high‑velocity categories (phones, wearables, gaming hardware).
- Surface provenance and timestamps in the UI so users can tell whether a recommendation is based on a live check or a cached snapshot.
- Harden normalization of SKUs and titles across merchants to prevent model priors from favoring older but well‑indexed models.
- Add explicit UI nudges for recency (e.g., “only show products released in the last 12 months”) and a visible “Provenance” toggle that lists the pages read. These are low‑friction controls that protect less technical shoppers.
Limits, caveats and unverifiable claims
Some claims repeated in testing deserve special caution. The assertion that “these assistants are training on product information that’s 18–24 months old” is plausible — model update cycles and training corpus windows can create multi‑month lags — but that precise 18–24 month figure is not a vendor‑published guarantee and cannot be independently verified without internal release notes or vendor confirmation. Treat specific numeric timelines about training/data age as
hypotheses unless providers make them explicit. Vendors typically publish rollout dates for product features (for example, OpenAI’s Shopping Research launch on November 24, 2025), but they do not disclose granular training‑corpus cutoffs for every product vertical.
Bottom line: practical verdict for the holiday season
AI shopping assistants are a meaningful UX advance: they lower friction, summarize large review sets, and centralize price‑tracking. When the retrieval layer is fresh and coverage is broad, they save time and yield high‑quality shortlists. However, the current reality is patchwork: varying merchant coverage, index freshness, and normalization shortcomings frequently produce
outdated top picks unless the user explicitly requests “current” models or cross‑validates.
For now, these assistants are best used as a rapid research assistant — a way to prune options and learn trade‑offs — rather than as a single source of truth for purchase decisions. Consumers should adopt a simple ritual: ask for the latest model, check the vendor page for the release date, and confirm price/stock at checkout. Vendors and platforms must accelerate API partnerships and add provenance labels; regulators should insist on disclosure for paid placements and allowlisting. If these changes happen, the assistants will fulfill their promise to make shopping smarter and faster. Until then, AI helps you
shop smarter — but not yet
shop blindly.
The holiday season will be a stress test: assistants will earn trust not by elegant prose but by up‑to‑the‑minute facts, transparent provenance and predictable coverage across the retailers people actually use. The technology is close; the data plumbing and transparency practices need to catch up.
Source: The Tech Buzz
https://www.techbuzz.ai/articles/ai-shopping-assistants-recommend-outdated-products/