Perplexity Computer: falsify a funded $30k side-by-side, not "agents replaced agencies"

Perplexity's customer story says the company invited Get Maine Lobster into a pilot to test whether its agent platform Computer could outperform the shop's previous manual campaign process. Get Maine Lobster was founded by Mark Murrell in 2009 and ships Maine lobster and wild seafood across the United States. The page puts average order value around $250 and says growth depends on email and paid social, with Meta ads described as essential but a beast to manage.

The mechanism is not a model taking over the company. It is a pipeline with a human gate. Computer was connected to Shopify, Attentive, Wicked Reports, GA4, and Xero. Each day it analyzes Meta and Shopify data, looks at existing Meta ads across ecommerce to form ideas, writes briefs, makes static ads, and sends them to Slack for Murrell to review. Video ads are sent to Higgsfield AI. Perplexity provided a $30,000 test budget, with the aim of meeting the customer acquisition cost (CAC) and return on ad spend (ROAS) that Murrell's manual campaigns usually reached. The published result is against "typical non-AI generated ads": Computer's side spent 2.1 times as much, brought in 3.5 times the revenue, and had 63% higher ROAS. Micro-testing creatives with different avatars, outfits, settings, and colors found winning ads with 14% lower CAC.

The test should sit on those ratios and on their limits: who paid for the trial, who approved the ads, and what the control was. It should not sit on an inference that AI has replaced ad agencies. The page names no agency and does not publish how much the control spent or sold.

Hard facts (Perplexity customer story, "How Get Maine Lobster drove 3.5x more ad revenue with Perplexity Computer", accessed 2026-11-06, Asia/Shanghai):

• Subject: Get Maine Lobster, founded in 2009 by founder and CEO Mark Murrell. The page says the New York Times, Good Housekeeping, and BuzzFeed have written about the company, and puts average order value around $250.
• Origin of the pilot: Perplexity invited the shop into a pilot and provided a $30,000 test budget to see whether Computer could meet the CAC and ROAS targets of Murrell's manual campaigns.
• Connected tools: Shopify, Attentive, Wicked Reports, GA4, and Xero.
• Daily loop: each morning Computer analyzes Meta and Shopify data piped into a "command center"; it analyzes existing Meta ads across ecommerce, not limited to direct competitors; through the day it generates ad ideas, writes creative briefs, makes static ads, and sends them to Slack for review; video ads are produced via Higgsfield AI.
• Human gate: Murrell's feedback is saved as a rule for later ads. The page says that over the first seven-day campaign, his approval rate rose from 20% to 90% with no feedback at the higher rate. Murrell is quoted as saying Computer does it all without a designer, researcher, or ads specialist, and that he only has to approve campaigns in Slack for them to ship.
• Reporting and scaling: Computer sends a daily PDF performance report to Slack. The example given for scaling winners is asking Perplexity, "can you scale all my winning ads by 20%?"
• Comparison result: the shop ran the Computer-generated Meta campaign alongside its typical non-AI generated ads. Computer's side had 63% higher ROAS while spending 2.1 times as much, bringing in 3.5 times the revenue. Micro-testing creatives with different avatars, outfits, settings, and colors found winning ads with 14% lower CAC.
• Fulfillment: the page says Computer also staggers shipping dates around the fishing schedule and updates arrival dates when purchases outstrip supply. Murrell says that at this rate the company will hit its fishing allotment, and that it is expanding into dinner kits, lobster rolls, and partnerships with local restaurants.
• Not on the page: start and end dates for the side-by-side (only "the first seven-day campaign" and results "in days"), absolute spend and revenue for the control, whether audiences overlapped, or any named ad agency. The page also does not publish a public price or availability plan for Computer.

Evidence strength: a single customer story written by the vendor, with a test budget supplied by the vendor. The $30,000 figure, 2.1x spend, 3.5x revenue, 63% higher ROAS, 14% lower CAC, and the rise in approval from 20% to 90% over seven days are all company statements on the page. No raw dashboard exports, third-party audit, or second customer replication is shown.

By arithmetic, 3.5 times the revenue on 2.1 times the spend is about 1.67 times the revenue per ad dollar, close to the stated 63% higher ROAS (which would be about 1.63 times). The gap may come from rounding or from different windows, but the page gives no absolute numbers that would let a reader check. Without control spend, a reader also cannot tell how much of the 2.1x spend is simply "more money in, more money out."

The check is: if the company or Perplexity publishes control spend, revenue, flight dates, and audience rules, outsiders can reconcile the 63% figure with the 3.5 / 2.1 ratio. If Computer is later sold to merchants who did not receive vendor funding, whether a second customer's ROAS comes close to these numbers would be stronger evidence than this page. Until then, the piece should be read as a single-shop pilot with human approval and a vendor-funded budget, not as proof that an industry has been replaced.

Steelman: if the Slack approval loop is real, and if Computer's side did show about 60% higher ROAS on the same accounting basis, that is a time-saving production path for a small merchant who lacks designers and media buyers. Even then, it does not prove that agencies have been replaced.

Strongest steelman: Murrell's pain is described in concrete terms. He had to jump between Shopify, Attentive, and GA4, and Meta ads required constant new creative and A/B tests, which took time he would rather spend on customer relationships. A loop that writes briefs, makes stills, ships after Slack approval, and sends a daily PDF report would be worth a pilot even without a 63% ROAS lift. The page says he can now spend more time on the boat, which matches the claim that repetitive work was handed off.

Steelman risk: the test budget came from Perplexity, and the page says the Perplexity team helped build the command center. That is a weak basis for treating the pilot as what an ordinary merchant would get when paying out of pocket, or as a blind test against a professional agency. The control is the shop's typical non-AI ads, not a named agency. An approval rate that starts at 20% and later reaches 90% shows that early outputs were not automatically usable; a person remained the quality gate. The "no feedback" 90% figure is also the page's own account. The creative micro-tests used different avatars and outfits; the page does not discuss licensing or disclosure for generated people or model images.

Separately, 3.5x revenue and 2.1x spend can be true at the same time, because spending more can raise revenue on its own. ROAS is the figure that says whether each ad dollar went further. It is written as 63% higher, but without absolute control numbers an outsider cannot run an independent check.

For small merchants, the next six months are about whether Computer publishes a price and a self-serve path, and whether a second public case appears without vendor funding. For ad agencies, this page is not industry evidence that agencies have been replaced. It is evidence that clients will start asking about a loop that ships ads after a Slack approval. For platforms and regulators, the source of generated creative, licensing for models or avatars, and how vendor-funded pilots are disclosed are likely to become contested sooner than any single ROAS number.

Anyone who quotes this story should keep three facts attached: Perplexity supplied the $30,000, the control was the shop's own non-AI ads, and the founder approved work in Slack. Drop those three and the numbers look larger than what the page actually states.

This Perplexity customer story should be judged on whether the spend, revenue, ROAS, and CAC figures versus the shop's usual non-AI ads still hold after fuller raw data are disclosed, under a $30,000 vendor-funded budget and Slack approval by the founder. It should not be judged as agents having replaced ad agencies. If the ratios hold and the method is published, it is a small-merchant advertising pilot worth discussing. If they do not hold, or if the method stays unpublished, it remains a single-shop case written by the vendor.

【No.2】