Verification traps on Amazon listings

Things on an Amazon page that read as authoritative and are not. Each of these produced a confident wrong answer at least once.

The status code

Amazon's bot wall returns HTTP 200 with a captcha body. Check the body, not the status. See fetch-routes.md.

The delivery date belongs to a ZIP you may not have set

An anonymous fetch always renders a delivery promise, for whatever ZIP the requesting IP resolves to. It never errors and never says which. Read ship_to before quoting delivery.

Also: a listing offers up to two delivery options, differing in cost as well as speed, and #deliveryBlockMessage returns only the first. The second one is where the Prime date usually is — including on an anonymous fetch. Read delivery_options, not delivery; contract in delivery.md.

Item Dimensions may be folded or unfolded, in the same field

Amazon's Item Dimensions L x W x H is populated by the seller, and for anything that collapses, sellers disagree about which state to publish. Two carts in the same category:

• 15.6" x 10.6" x 3.2" — folded
• 17.52" x 10.23" x 37" — erect, where 37 in is the handle height

Nothing in the field distinguishes them. Take folded size from the bullets, where sellers advertise it deliberately, and treat a large height in the spec table as suspect.

The same conflation runs through platform trucks and hand carts generally: a height quoted in the title is almost always handle height, not deck height. If a "1 m platform truck" matters, verify the deck.

The spec table and the bullets disagree, on the same listing

One cart's first bullet opens "At just 4 lbs" while its own spec table says Item Weight: 2.8 pounds. Neither is flagged; both are on the page. Where the number decides the purchase, read both and record the disagreement rather than picking the one you prefer.

The first price on the page is the List Price

The struck-through "List Price" is rendered in the same block as the price you would pay, and on discounted listings it comes first. An extractor that takes the first price span therefore reports the list price, silently, with no marker.

Measured 2026-08-19 on ASIN 0140449132: the plugin reported $16.00 against a real $12.91 — a 24% overstatement. Skip any price whose enclosing a-price span carries a-text-price, data-a-strike="true" or basisprice.

#corePrice_feature_div is additionally empty on books; the price lives in #corePriceDisplay_desktop_feature_div.

And the raw HTML and the live DOM disagree about where the value is. In the HTML the priceToPay screen-reader span carries it; in the rendered DOM that span is blank and the price is assembled from .a-price-symbol / -whole / -fraction. A selector proven against one will return nothing against the other.

A price with no buy box belongs to something else

A variant parent with no featured offer renders no buy box, no delivery block and no seller row — but the page still carries prices belonging to other variants, and an extractor will return one.

Measured 2026-08-19 on B003J9LZE4: $108.99, availability: "", no delivery. Nothing said the item could not be bought. Note this defeats the usual test: Currently unavailable appears only inside JavaScript config strings on such a page, never as rendered text.

The reliable check is the absence of #add-to-cart-button and #buy-now-button. amazon_fetch.py reports it as buyable and attaches a no_offer_warning.

A real browser can disagree here: navigating to the same parent URL redirected to ?th=1 and pinned a variant, so the browser saw a buy box where the anonymous fetch saw none. Same ASIN, two answers, both correct for what they fetched.

The stock line has a second sentence

In stock and In stock. Usually ships within 4 to 5 days are different answers, and reading only the first span of #availability returns the same text for both. The handling-time sentence is often the reason a September date sits under an in-stock badge. Only 17 left in stock - order soon lives in a different colour class again.

textContent on a fetched document pulls in scripts

DOMParser builds a document with no layout, so innerText is empty and textContent is the obvious substitute — but it concatenates the contents of every <script> and <style> too. Three consequences, all observed:

• #availability returns In Stock followed by an inline stylesheet;
• #olp_feature_div returns nothing but CSS;
• the page's inline JSON — including session identifiers — lands in your context, and the Chrome extension blocks the whole tool result as cookie/query-string data.

Strip script, style and noscript before reading text from a parsed document.

?aod=1 returns 200, 3 MB, and no offers

The "all sellers and offers" panel is client-rendered. Fetching /dp/<ASIN>?aod=1 or /gp/offer-listing/<ASIN> same-origin returns HTTP 200 and a full-size page containing zero #aod-offer nodes — a textbook success-shaped failure. The older ajax endpoints (/gp/product/ajax?...experienceId=aodAjaxMain and variants) now return 404.

Opening ?aod=1 in a real tab does work, for a human. Treat it as a handoff URL, not a data source.

Sold by and Shipper / Seller are different facts

#sellerProfileTriggerId and the merchantName JSON key are both gone from the page as of 2026-08-19, so any extractor keyed on them returns null on every listing while looking like it simply found no seller.

What renders today is #merchantInfoFeature_feature_div, and its label is itself the fulfilment signal:

Label	Means
Sold by X + #fulfillerInfoFeature_feature_div = Ships from Amazon	FBA — merchant sells, Amazon fulfils
Shipper / Seller X, fulfiller div empty	The merchant does both

Both values are duplicated in the markup for the responsive layouts (Sold by waveshare waveshare Sold by waveshare). The repeat is a layout artefact, not a second seller.

A dead ASIN 404s on fetch but not on open

A same-origin fetch of /dp/<ASIN> for an ASIN with no listing on this storefront returns a genuine HTTP 404 with a ~2 KB body. That is a real existence check, and it is one the xdg-open route cannot perform — see An opened tab is not a live product above. Two routes, and only one of them can tell you whether the thing exists.

Grid prices are positional; listing prices are not

A price in a search grid is the price nearest the link in a flattened page, not a value read from a labelled field. amazon_fetch.py search pairs each result with the price inside that result's own data-asin block, which is structural and better — but it is still a shortlisting tool.

Open the listing before recommending anything. Three facts exist only there: who actually sells it, whether it is Currently unavailable (which never appears in search results), and how many competing offers there are at what price.

Pack sizes are separate ASINs with separate everything

Pack pricing is routinely non-monotonic — a 4-pack can be cheaper per unit and arrive sooner than the 2-pack of the identical item. Price every pack size separately and write out the per-unit figure.

Ratings hide the failure mode

A low rating on a low review count is the highest-value signal available here. Established brand names ship rebadged product more often than the name suggests, and an unbranded listing with 2,000 ratings is not thereby trustworthy.

Read at least one critical review on anything expensive. The average hides the failure mode; the review text names it.

Review authenticity cannot be checked programmatically — Fakespot shut down in 2025 and nothing replaced it, and AI-written reviews read cleanly enough that linguistic tells no longer discriminate. Use structural signals instead: who the seller is, whether the brand exists off Amazon, whether Date First Available is recent. Say the review score is unverified rather than implying otherwise.

Sponsored rows are not results

Filter sponsored placements out before any judgement. They skew heavily toward unknown brands buying placement. amazon_fetch.py marks them sponsored: true.

A filter that prunes to nothing is a finding

An empty filtered result set is a fact about the category or the address, not a successful search. Say so rather than returning an empty table, and say what the filter cost — the result count before and after.

Likewise, report rows with no price as "no current offer" rather than dropping them. A missing price is not a cheap price.

An opened tab is not a live product

xdg-open hands a URL to an already-running browser and exits 0 without ever seeing the response. A retired, mistyped or region-wrong ASIN opens Amazon's "Sorry, we couldn't find that page" placeholder with exactly the same exit status as a live listing. Opening is a handoff, not a check.

The region case is the frequent one: an ASIN copied from amazon.co.uk, amazon.de or amazon.ca often has no .com listing at all. Same identifier, different catalogue. Rewriting the domain produces a URL that looks right and resolves to nothing.

Also note that an ASIN is ten characters of [A-Z0-9] and not necessarily B0-prefixed — books keep their ISBN-10, including trailing-X forms. A validator that insists on B0 rejects valid input.

What no route can tell you

• Price history. No free programmatic source. Keepa is the working paid one.
• Listing age, when Date First Available is absent — which is common on electronics. Do not substitute review count for age.
• Per-card brand. The brand row renders on brand-navigational queries only (17 of 21 cards on one query, 0 of 20 on another). The facet rail is the source of truth; never build a brand list from cards.