Top 10 AI Agents forEcommerce Brands in 2026

Every ranking of ecommerce AI agents you will read this year is built on a number that the vendor invented, defined and invoiced. Open five of them and you will find the same shape: this agent resolves 70% of tickets, that one deflects 60%, a third contains 85%. Those are printed as though they were measurements taken by someone. They are not. They are counts of a unit each seller defines for itself, published by the seller, with no methodology and no neutral party checking a single one.

The claim under test here is narrow: can a DTC operator rely on any published performance figure in this category? After checking every vendor page in this roster on 23 August 2026, our answer is no — and the reason is not that vendors are lying. It is structural. Zendesk's own support documentation explains that an automated resolution is counted after 72 hours of inactivity, subject to an LLM verifying the reply was relevant, where the shopper either gave positive feedback or gave no feedback at all, and where no human agent replied. Read that as an operator rather than as a buyer of software. A shopper who asked a question, got an answer that did not help, gave up and bought elsewhere leaves no feedback and generates no human reply. Seventy-two hours later, that abandonment is counted as a resolution and billed to you as one.

It gets worse before it gets better, because the vendors in this market do not even bill in the same kind of unit. Across the thirteen products we examined we found five mutually incompatible billing units — per resolution, per seat, per order, per credit and per share of your GMV. A buyer comparing $0.90 to $0.99 to $2.00 is comparing three definitions written by three different sellers, at least one of which includes a $750 monthly floor before the first ticket is touched. That is not a price comparison. It is a category error with a spreadsheet around it.

So this ranking does something narrower and more useful. It scores only attributes you can re-check yourself in under an hour with a browser, it records the unit in every pricing cell rather than just the number, and it names what it refused to score and why. This piece sits inside a wider set of vertical rankings built on the same method; the cross-industry version is our guide to the top AI agents in 2026. What follows is the ecommerce edition, written for a brand somewhere between $2M and $100M in revenue, not for an enterprise retailer with a procurement department.

One piece of context worth having in front of you while you read: the market is not shrinking, and nobody needs to sell you a crisis to justify the spend. The US Census Bureau's Quarterly Retail E-Commerce Sales report for the second quarter of 2026, released 18 August 2026, puts adjusted US retail e-commerce at $340.2 billion, 17.1% of total retail sales, growing 12.2% year over year against 6.7% for retail overall. That is a real, dated, primary figure from a statistical agency, and it is the only kind of number in this article you did not have to take from a seller.

We start with the refusal, because it is the more important half.A resolution is a billing definition set by the seller, not a performance metric — so resolution rate, deflection rate, containment rate, automation rate, CSAT lift, hours saved and ROI are all excluded from the scoring. Not because they are unimportant, but because there is no neutral instrument measuring any of them, and printing two vendors' self-reported rates side by side implies a common standard that does not exist.

The evidence for that refusal is not an opinion we formed. It is on the vendors' own pages. Zendesk's automated-resolutions documentation states that a resolution is counted after 72 hours of inactivity where the end user gave positive feedback or gave none, provided no human agent responded, and that each resolution is verified by an LLM rather than by a person. Gorgias's published pricing page uses a comparable 72-hour window to decide whether a later human touch is re-billed. Both are defensible engineering choices. Neither is a measurement of whether a shopper was helped, and neither vendor claims it is — that claim gets added later, by marketing, and then by listicles.

On the independent-evidence question, we want to be precise rather than sweeping. No independent benchmark of these commercial products exists — we looked and located none as of 23 August 2026. That is a narrow claim. Independent evaluation does exist elsewhere in AI agents; it just measures models or modalities rather than the specific tools you would actually buy. In ecommerce specifically, there is nothing: no head-to-head, no third-party accuracy measurement, no standard definition of the unit being counted.

What does exist is the one occasion a US regulator audited a vendor's published AI performance metrics, and it is worth knowing what happened. In In the Matter of Presto Automation Inc., Securities Act Release No. 11352, Administrative Proceeding File No. 3-22413, entered 14 January 2025, the SEC addressed a company that had told investors its drive-thru voice AI delivered “over 94% accuracy even in noisy environments” and 95%–99% automated order completion. The Commission found the product “lacked the capability to take orders on their own and required substantial human involvement,” with “human order takers located abroad (primarily in the Philippines and India), who processed the vast majority of drive-thru orders.”

Three framing points are mandatory and we will not skip them. Presto consented to the order without admitting or denying the findings. The remedy was a cease-and-desist order with no civil penalty — it was not a fine. And the findings are against Presto only; the order's “Supplier A” is a different company against whom the SEC made no findings whatsoever. It is a restaurant technology case, not an ecommerce one. We cite it for a single, transferable proposition: this is what happened the one time a regulator with subpoena power checked a vendor's published AI performance metric, which is why we score only what a buyer can verify.

The single most useful thing you can do before comparing any two vendors in this market is write down the unit next to the number. We found five kinds of unit across thirteen products, and they are not variations on a theme — they are five different theories of what you are buying. You are buying an outcome, or a seat, or an order, or a credit, or a share of your revenue. Those are not comparable quantities, and no amount of spreadsheet formatting makes them one.

The cleanest demonstration is two vendors publishing the same headline figure. Gorgias publishes $0.90 per AI Agent resolution on annual plans. Siena publishes $0.90 per automated ticket. Identical numbers. But Siena's structure adds a $750 per month platform fee before the first ticket. At 300 automated tickets a month, that is $1,020 against $270 — a 3.8× difference hidden entirely inside the unit. At 3,000 tickets a month the gap narrows to $3,450 against $2,700, and by 10,000 the platform fee has become a rounding error. Neither vendor is being deceptive. The headline figure is simply not the price.

Then there is the structure nobody talks about. Triple Whale prices on GMV bands — free, $219 a month, and a tier shown at $1,290 a month, with bands running from “up to $2M” through “$7M+”. Read that as an operator: the vendor's revenue rises with yours whether or not you used the agent more. For a brand between $2M and $100M in revenue, that is the single most consequential pricing structure in this roster, because it is the only one where a good year automatically costs you more software.

And exactly one vendor has removed its own incentive to count generously. Richpanel publishes a flat monthly automation tier at $120 and $360 with unlimited resolutions, positioned explicitly against per-resolution competitors. Whether the product suits you is a separate question we have no independent evidence to answer. But structurally, a flat rate is the only arrangement here where the vendor gains nothing from a liberal definition of the word being counted — and that is a checkable, scoreable property, which is precisely the sort of thing this method rewards.

Every cell below is either a citation you can open or the literal words “not publicly disclosed” / “not verified”. Nothing is estimated. A low position is not a verdict on product quality — it means less was verifiable from a browser on 23 August 2026, which is the only thing this method claims to measure.

Two observations about the shape of that table are worth more than any individual row. First, compliance posture could not be verified for 7 of the 13 vendors we examined, and for two of them the security URL returned a 404 — there is no trust page to read. Second, ownership of record is unverified for four of them, in a market where an acquisition happened during our research window. Both facts are findings, not gaps in our work, and we score them as findings.

Three products are widely listed in this category and should not be. Naming them is more useful than adding an eleventh entry, because a stale roster is the failure mode that gets a ranking dismissed by the operator it was written for.

• Wonderment — acquired by Loop Returns. Its site now serves a page titled “Wonderment is now Loop Tracking!”, with a canonical URL pointing at loopreturns.com and a banner announcing the acquisition. Ranking it as an independent post-purchase vendor in 2026 is a factual error. We could not establish the acquisition date, so we do not print one.
• “Zoe by Gorgias” — no such product exists on Gorgias's own pricing page, which names Starter, Basic, Pro, Advanced and AI Agent. Under this method, a product we cannot confirm exists in 2026 does not get listed.
• Daydream — a real, trading company, but an SEO and AI-search visibility firm combining agents with dedicated human experts, whose named clients are B2B and technology businesses. That is an agency-plus-software service, not a DTC ecommerce agent product. Out of scope, not defunct.

Three further vendors are real, trading, and relevant, but did not rank because too little was verifiable. Rep AI has the best-documented corporate record in the entire roster — a $6.2 million strategic follow-on round announced 28 May 2026, led by Silicon Road Ventures with Osage Venture Partners, Flashpoint Venture Capital and Zendesk as a strategic investor, following an $8.2M Series A in August 2024. That last detail is worth holding: Zendesk invests in Rep AI, which is legitimate context when you are comparing the two. But Rep AI's pricing page contains no dollar figures at all— only a trial offer and a “5× ROI guarantee” — and hellorep.ai/security returned a 404. Its Helpdesk is labelled beta on its own site.

Narvar markets agentic post-purchase products and has the broadest documented integration set we found — Shopify, Zendesk, BigCommerce, Klaviyo, Magento, Gorgias, Salesforce and Attentive named on its own site — but publishes no pricing and routes to a demo. Its homepage also carries four striking statistics about package theft, delivery-issue costs and consumer trust, none of which is sourced on the page, so none appears in this article. It is enterprise-weighted, and for a $2M–$100M brand that is a fit question before it is a capability question. Kustomer is trading and positions as an AI customer service platform and CRM, but its pricing page is fully client-rendered and returned nothing to us, and — more importantly — we could not establish who owns it in 2026 from public record. It has been acquired and divested before. We do not state ownership we have not verified, and an unresolved owner is a genuine procurement risk, not a footnote.

Two statistics dominate ecommerce AI sales conversations and neither should travel in the form it usually arrives. Refusing them is not pedantry — a vendor who opens with an unsourceable number has told you something about how the rest of the deck was assembled.

“About 70% of carts are abandoned.” The figure is real and its provenance is public. Baymard Institute publishes 70.22% and states on the same page that the value is an average calculated from 50 different studies, gathered so people could cite one number instead of fifty. The inputs span 2006 to 2025 — the oldest is a 2006 study at 59.80%, the newest a 2025 one at 71.72% — and the range across the set runs from 55.00% to 84.27%. Most of the sources are vendors selling abandonment remedies. Baymard itself adds the caveat almost nobody quotes: a large portion of cart abandonments are a natural consequence of how users browse ecommerce sites, and are largely unavoidable.

Baymard's own primary research is the citable part, and it is more useful anyway: 42% of US online shoppers have abandoned a cart because they were just browsing or not ready to buy. Excluding that group, the leading causes are extra costs too high (40%), delivery too slow (20%), not trusting the site with card details (19%), forced account creation (18%), and checkout too long or complicated (17%). Baymard also finds the average US checkout shows 23.48 form elements against an ideal of 12 to 14.

“It costs five times more to acquire a customer than to retain one.”We went looking for the origin and could not establish one. The attribution shifts between at least three different sources, and the multiplier quoted in the wild ranges from five to twenty-five depending on who is repeating it. Under this article's own standard, an untraced statistic is refused — so we do not print it, in any multiplier, attributed or not. If a vendor quotes it at you, ask where it comes from and watch what happens.

The same refusal applies to every deflection, containment, automation and resolution rate in this market, and to the cluster of pilot-failure statistics that circulate as ambient dread — “95% of GenAI pilots fail” describes 52 interviews and 153 conference surveys about custom-built tools, not a purchase decision about a commercial ecommerce agent. Fear priced from a misread statistic is still a misread statistic.

The constraint that decides whether any of this works in ecommerce is not model quality. It is that three of the most commonly automated workflows — reviews, cancellation and personalisation — each sit directly on top of a rule with penalties attached, and the market is confidently wrong about all three.

Start with reviews, because a vendor in this very roster ships a Reviews Agent. The FTC Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465, took effect 21 October 2024 and is in force and unamended as of 23 August 2026 — we confirmed that by pulling Part 465 live from the eCFR and by searching the Federal Register for any amending or vacating document since September 2024, which returns nothing.

And now the trap that most published summaries fall into: the rule does not ban review hijacking. Proposed § 465.3, which would have addressed reusing a review written for a substantially different product, was never finalised. The Commission decided not to proceed with it, dropped the associated definition, and the section is codified [Reserved]. An article — or a vendor datasheet — that says the rule bans review hijacking has printed a false statement of law.

What it does reach is fake reviews and testimonials, buying reviews the business knew or should have known misrepresent that the reviewer exists or used the product, incentives conditioned on sentiment, undisclosed insider reviews from officers, managers, employees or their relatives, company-controlled sites posing as independent, review suppression through groundless legal threats or rating-based filtering, and fake indicators of social media influence. The operative text never uses the word “AI.” AI coverage comes from footnote 35 of 89 FR 68034, where the Commission wrote that AI tools make it easier for bad actors to pollute the review ecosystem by generating large numbers of realistic but fake reviews, and that AI-generated reviews are covered by the final rule. Cite it that way or not at all.

The maximum civil penalty for a violation of a section 18 trade regulation rule is $53,088. There is no 2026 adjustment — we searched for one and none exists. If a vendor quotes a higher, more current-sounding figure, ask which document sets it.

— 16 CFR 1.98(d), as amended at 90 FR 5581 (17 January 2025)

The two provisions an automated system trips first are § 465.5 and § 465.7. An agent that mass-solicits reviews from staff, or from anyone with an undisclosed material relationship, is inside the insider-review provisions. An agent that routes negative reviews away from display — or applies a “quality” filter that happens to key on star rating — is inside the suppression provisions, and the carve-outs there are explicitly for sentiment-neutral criteria such as defamatory content, personal information or off-topic posts. A sentiment filter is the opposite of sentiment-neutral. This is the hardest human-in-the-loop line in the article, and it should be enforced by removing the capability rather than by instructing the model not to use it.

Where does that bite in practice? Anywhere reviews are collected and displayed on your own storefront. Our Grillz custom jewellery build is a Shopify Plus store whose delivered feature set includes a customer review system with photos alongside Klaviyo email and an Instagram feed — an ordinary, well-built DTC review surface. The moment an agent is given write or moderation authority over a surface like that, Part 465 is the governing document, and the audit log of what was suppressed and on what stated ground is the artefact you will be asked for.

For a subscription brand the practical translation is short. Do not let a vendor sell you a compliance feature against a rule that no longer exists — and do not read “vacated” as permission to make cancellation harder. If you ship to California, § 17602 also requires renewal reminder notices (3 to 21 days before a trial or promotional period of over 31 days ends, and 15 to 45 days before an annual-or-longer term renews), a retainable acknowledgment containing the cancellation policy, and retention of proof of consent for three years. Other states have their own automatic-renewal statutes with different notice triggers; we verified California only, so we say “state automatic-renewal laws, of which California's is the strictest we verified” and leave the enumeration to your counsel.

On privacy, the burden is widely misplaced. The CPPA's CCPA updates, cybersecurity audit, risk assessment and ADMT regulations took effect 1 January 2026, and CCR § 7200(b) sets ADMT compliance for significant decisions at 1 January 2027. But “significant decision” is defined at CCR § 7001(ddd) as one resulting in the provision or denial of financial or lending services, housing, education enrolment or opportunities, employment or independent contracting opportunities or compensation, or healthcare services. Product recommendations, merchandising, marketing personalisation and support triage are not significant decisions, and behavioural advertising did not survive into the final scope. The one edge worth flagging rather than asserting: § 7001(ddd)(1) includes installment payment plans within financial or lending services, so an agent deciding pay-over-time eligibility may be in scope.

What binds a personalisation agent today is the Global Privacy Control, under CCR § 7025(b) — mandatory now, not in 2027. The regulation provides that a business which sells or shares personal information shall process a conforming opt-out preference signal as a valid request to opt out of sale or sharing, and names an HTTP header field or JavaScript object as an example format. If your recommendation, retargeting or lifecycle agent is fed by data that is sold or shared under the CCPA, an inbound GPC signal must switch that off for that visitor. It is testable in about a minute from a GPC-enabled browser, it is far more commonly breached than the ADMT rules are, and it is the first thing we check in a vendor security review.

And if you ship to Europe, one more live obligation. EU AI Act Article 50 transparency obligations applied from 2 August 2026 — three weeks before this was written. A person must be told they are interacting with an AI system unless that is obvious from the circumstances. It is a common and expensive mistake to believe the EU delayed the AI Act to 2027: Regulation (EU) 2026/1744, the Digital Omnibus, entered into force 27 July 2026 and deferred only the high-risk obligations — Annex III standalone systems to 2 December 2027 and Annex I embedded systems to 2 August 2028. Article 50 was not deferred. The remedy is a disclosure line rendered by your platform rather than generated by the model, because a model can be talked out of its own disclosure and a template cannot.

Finally, the constraint nobody has solved. Prompt injection is unsolved. Any agent reading untrusted input — customer email, review text, supplier documents, return-reason free text — is exposed to instructions embedded in that content. There is no vendor in this roster, or anywhere else, who has fixed it. Treat it as a blast-radius problem: the agent identity that reads external content should not hold the credential that moves money, changes prices, or publishes to a public surface. That separation survives a successful injection. Model quality does not.

This is an illustrative scenario, not a client result. Every figure below is either a published vendor price or arithmetic we performed on published vendor prices, and you can redo it with your own numbers in a spreadsheet. Consider a DTC brand doing roughly $18M a year, running about 2,400 support conversations a month, of which the agent handles 1,000 without a human reply.

1. 1.Gorgias: 1,000 AI resolutions at $0.90 on an annual plan = $900 per month, plus the ticket-based plan fee for the remaining 1,400 conversations and any automated-interaction overage at $1.50.
2. 2.Siena AI: $750 platform fee plus 1,000 automated tickets at $0.90 = $1,650 per month, before implementation, which is contact-sales.
3. 3.Intercom Fin: 1,000 outcomes at $0.99 = $990, plus seats — six seats on the $85 tier adds $510, for roughly $1,500 per month. Or run Fin on a third-party helpdesk at $990 with no seat cost, subject to the minimum commitment.
4. 4.Zendesk: 1,000 automated resolutions at the published pay-as-you-go rate of $2.00 = $2,000 per month, plus seats, with committed usage documented as cheaper than overage.
5. 5.Richpanel: the flat automation tier at $360 per month, marketed as unlimited resolutions, plus the per-user helpdesk and per-order-band self-service lines as configured.

The same 1,000 conversations cost somewhere between $360 and $2,000+ a month depending only on which theory of the unit you bought.That is a 5.5× spread before a single question about quality, and it is entirely visible on public pricing pages. Now layer the definitional problem on top: on Zendesk's published counting rule, some unknown share of those 1,000 “resolutions” are shoppers who never replied because the answer did not help. You are billed for those at $2.00 each, and they appear in the vendor's dashboard as successes.

Two of our own ecommerce builds are worth naming here for what they say about the surfaces an agent would touch, and it is worth being exact about what they were. The Players Choice CBD platform is a Shopify Plus store built with Recharge for subscriptions, age verification and compliance systems, lab result transparency, a certificate-of-analysis viewer, a loyalty programme and an affiliate system. It was a storefront and subscription-commerce build, not an AI agent deployment, so nothing about its reported outcomes transfers to an agent business case — and we will not pretend otherwise. What it does illustrate is the shape of the exposure: a brand with recurring billing and a compliance-sensitive category is exactly where ROSCA § 8403 and California § 17602(d) land, and where a retention agent inserted into the cancellation path would be the riskiest thing on the roadmap.

The honest conclusion from the arithmetic is unglamorous. Before you shortlist anything, export ninety days of your own helpdesk data, tag the top ten contact reasons, and find out how many of them are answerable from data you already hold. If order status and returns are 60% of your volume, the winning vendor is whichever one is cheapest at your volume in a unit you can predict — and that is a spreadsheet question, not a demo question.

Ecommerce agent deployments do not usually fail by hallucinating. They fail by drifting — in catalogue data, in billing units, in cancellation friction, and in permissions nobody re-checked after a tag manager change. Each row below pairs the failure with the signal that catches it early and the rollback that limits the damage.

The one that surprises operators most is the second row. A per-resolution or per-credit contract can grow faster than your ticket volume without anybody doing anything wrong, because the counting rule and your traffic mix both move. The defence is boring and it works: reconcile the vendor's resolution count against your own cost per contact every month, and put a written right to advance notice of pricing and model changes into the contract. Zendesk changing its automated resolution model on 18 May 2026 is the proof that this happens on the vendor's schedule, not yours.

Write these boundaries into configuration, not into a prompt.A prompt is a request; a tool permission is a constraint. Anywhere the consequence is money leaving, a price changing, or a public surface being written to, the boundary should be enforced by the absence of the capability rather than by the model's cooperation.

The reviews rows are the ones to enforce hardest, because they carry statutory penalties rather than commercial ones, and because they are the rows most likely to be crossed by a well-meaning default. Nobody sets out to build a sentiment filter. Someone adds a “hide reviews below three stars while we investigate” rule during a bad week, and § 465.7 does not care that it was temporary.

These are our standard bands, published so you can benchmark any proposal — ours or anyone else's — against a real number rather than a range invented on a call.

Senior-led work is $150–$225 per hour; ongoing retainers run $2,500–$9,500 per month. Every engagement carries a 30-day post-launch warranty, and full source-code and IP ownership transfers to you — there is no vendor lock on what we build. We deliver a fixed-price phased proposal within 5 business days of a discovery call. Frenchy Digital is a senior-led Black-owned agency in Los Angeles, reachable at +1 (424) 272-5601.

A note on where the money actually goes in an ecommerce build. It is rarely the model. It is the catalogue and order-data plumbing, the escalation design, the monetary ceilings and idempotency around anything that moves money, the GPC and disclosure work, and the instrumentation that lets you reconcile a vendor's counting against your own. Whether that work belongs inside a purchased platform or in something you own is the real decision, and it is the one we work through in our build-versus-buy analysis for AI agents.

• A resolution, deflection or containment rate presented as neutral fact, with no published definition of the unit being counted. Ask for the definition in writing before you ask for the number.
• A compliance badge whose certificate belongs to the hosting provider rather than the vendor's own legal entity. Confirm the certificate names the vendor, not its cloud.
• “SOC 2” without a type. Type I attests to design at a point in time; Type II attests to operating effectiveness over a period. One vendor here states “SOC2 certified” and does not say which.
• No trust page at all. Two vendors in this roster return 404 on their security URL. That is not an oversight to work around; it is the answer to your question.
• A pricing page with no dollar figures but a quantified ROI guarantee. The guarantee is a marketing claim; the missing price is the actual disclosure.
• A compliance feature sold against the FTC “click to cancel” rule. That rule was vacated in 2025 and struck from the CFR in February 2026. Selling against it in 2026 means the datasheet has not been read since it was written.
• Any claim that the reviews rule bans review hijacking. Proposed § 465.3 was never finalised and is codified [Reserved].
• Any claim that the EU delayed the AI Act to 2027. The Digital Omnibus deferred high-risk obligations only; Article 50 applied from 2 August 2026.
• A retention or save-offer flow marketed as a cancellation feature. ROSCA and California § 17602(d) both point the other way, and a regulator can walk your funnel as easily as a customer can.
• A model-training clause that lets the vendor use your customers' conversations to improve its own models. Strike it or narrow it, and get deletion schedules in writing before signature.
• A promise that prompt injection is handled. It is not handled anywhere by anyone. The right answer describes blast-radius controls and credential separation.
• An integration list that lives only on the vendor's marketing page and not in its documentation. Ask for the docs URL; if the connector is real, it has one.

This section is part of the product, not a disclaimer bolted to the end of it. Everything below is something we tried to check on 23 August 2026 and could not, or could only check partially. Naming it is the difference between a ranking and a sales page.

• No independent benchmark of these commercial products exists. We located none. Every performance figure in this market is vendor-published about the vendor's own product, using a unit that vendor defines.
• Compliance posture is unverified for 7 of the 13 vendors we examined. Siena AI and Rep AI have no locatable trust page at all — both security URLs returned 404. Kustomer, Richpanel, Octane AI, Triple Whale and Narvar were not fetched this session. Only Intercom (comprehensive), Gorgias (SOC 2 Type II) and Loop (SOC 2, type unstated) are verified.
• No PCI DSS attestation was located for any vendor here, including those that touch refunds, instant exchanges and store credit. Any PCI claim in a proposal needs a fresh check against the vendor's own attestation.
• Ownership is unverified for Kustomer, Zendesk, Narvar and Triple Whale. Kustomer is a known acquisition-trap name — it has been acquired and divested before, and we do not state a 2026 owner we did not confirm.
• Klaviyo and Zendesk pricing were geo-served non-US locales during our fetch, mixing currencies and locale-specific plans. We print no seat or profile figures for either, rather than printing numbers we did not read from a US context.
• Shopify's agentic-commerce surfaces — catalogue, knowledge base, universal cart — were not verified at all; help.shopify.com returned 403 to our fetcher. We make no claims about what has shipped or whether it is generally available, and neither should a ranking written before yours.
• The integration matrix is thin. Only Narvar (eight named platforms), Rep AI (Shopify, Klaviyo, WooCommerce, Salesforce) and Octane AI (Shopify, Klaviyo) had integrations verified from their own pages. BigCommerce, ShipStation and Recharge support was not verified for any vendor except Narvar's BigCommerce listing.
• The Wonderment acquisition date could not be established — the announcement URL returned 404. We say the product is now Loop Tracking and we do not date the transaction.
• Siena's product page marks some agents as coming soon and the rendering was ambiguous about which. We do not claim the Shopping Agent or QA Agent are generally available.
• Gorgias's tier-to-price mapping was extracted from structured data on its pricing page rather than from the rendered layout. Re-read the rendered page before you bind a specific number to a specific named plan in a contract.
• Allegations circulating in third-party blogs about systematic double-billing by one vendor were not verified against that vendor's documentation, so they do not appear here as anything other than this sentence.
• Card-network liability allocation for card-not-present fraud, and the 3-D Secure liability shift, are governed by Visa and Mastercard operating rules — private contracts we did not obtain. We state the principle and tell you to confirm your position with your acquirer.
• We verified California's automatic-renewal law only. Other states have their own statutes with different notice triggers, and we deliberately do not enumerate them.
• The FTC's Consumer Sentinel 2025 Data Book is not published at the FTC's standard URL as of 23 August 2026 — it returns 404 and the newest indexed edition is 2024. No 2025 Sentinel figure appears in this article.
• The FTC's July 2026 proposed policy statement on suppression of accuracy in AI systems (91 FR 41638) is a proposed policy statement, not an adopted rule, and it is tangential to ecommerce agents. It governs nothing today and we rely on it for nothing.
• Vendor customer counts, integration counts and funding figures are as published by the vendor or reported in a dated press release, and we did not independently audit any of them.