Market research with an LLM:
is there room in the standing-desk market?

Is there room for a new standing desk in the US, and if so, where? It's a typical early question for a product team or an investor: crowded category, lots of opinions, no budget yet for a research agency. A good answer needs four things: how much people want it, what is selling and at what price, what buyers complain about, and what they ask before they buy.

The five desks for reviews covered the main types on the list: a manual converter, a small electric desk, a 63-inch electric desk, a desk with drawers and a mobile desk. Each tool returns JSON, so the whole collection step is five API calls; it took about three minutes, most of it waiting for Amazon.

We gave the model three jobs, and kept everything that is counting for code.

1. Tag every review. claude-sonnet-5.5 read all 100 reviews in one call and gave each up to three themes from a fixed list (wobble, motor failure, easy assembly, damaged parts and so on) plus one short quote copied from the review. Code then checked that every quote really appears, word for word, in its review: 7 of 100 didn't, so we dropped them.
2. Read the Reddit posts. A second call grouped the 27 posts into what buyers are trying to decide, the problems they report and the setups they have, each with the ids of the posts behind it.
3. Write the brief. The last call got only the numbers that code had computed (theme counts, price medians, the trend by month) and the verified quotes, with one rule: every claim must cite its number.

The model never counted anything itself. Letting it tally themes across 100 reviews invites confident, wrong totals; asking it to tag and letting code count does not.

Nearly every best-selling desk is a standing desk. 45 of the top 50 in Amazon's Office Desks & Workstations are standing or height-adjustable. By type: 12 mobile or rolling desks, 11 converters that sit on an existing desk, 10 electric desks, 8 L- or U-shaped electric desks and 4 frames sold without a top.

It is a cheap category. Median prices were $61 for mobile desks, $142 for frames, $150 for electric desks, $160 for converters and $215 for L-shaped desks. Of the 44 standing products with a listed price, 14 cost under $100 and 21 cost $100–199; only 9 cost $200 or more.

Demand jumped this spring, then mostly settled. US search interest averaged 16 to 18 every year from 2022 to 2025. In April 2026 it averaged 80, peaking at 100 in the week of 12 April, then fell back. From July to October 2026 it averaged 22, about 35% above the same months of 2025.

Bad reviews are about things breaking. 28 of the 100 reviews gave one or two stars. The themes behind them:

In their own words: "Motor stopped working after not even nine months", "Product arrived with cracked desktop", "the drawers is plastic when it is stated and advertised to be metal". The 62 four- and five-star reviews praised easy assembly (31), build quality (28) and simply that it works (23).

On Reddit, wobble dominates. 8 of the 27 posts are about a desk that shakes at standing height, usually under monitors or speakers. The other recurring questions: is it worth the money (4 posts), which model fits my budget (6), and how big and thick a top do I need (4).

Its three opportunities, lightly shortened:

1. A desk that doesn't wobble under a heavy setup. Wobble appears in 8 of 27 posts and 6 of 28 low reviews, and 3 posts ask whether a desk stays stable with several monitors. Test: film the same load on 2–3 desks from the $100–199 band and measure waitlist signups from the footage.

2. Reliability backed by a real warranty. Motor failures are 7 of 28 low reviews and customer service 6 of 28; two posts flag dual-motor frames with one year of cover. Test: run two ads on a mock listing, with and without a longer warranty, and compare click-through.

3. A clearly specified mid-price desk. Only 9 of 44 standing products cost $200 or more, and buyers say they can't tell cheap from expensive. Test: publish a comparison of motor type, top thickness, load and warranty, and measure email capture.

It also listed what the data can't say: no sales volumes or margins, a relative search index rather than demand in units, reviews from only five products, and too few Reddit posts to generalise from. That section is worth as much as the opportunities.

• The "doubling" that wasn't. The first summary said 2026 interest was double 2025's, which is true of the yearly average and misleading: the whole difference is a spring spike. Looking at the months before writing anything fixed it. Always ask for the shape, not just the average.
• Product types by keyword. Our first pass called anything with "frame" in its title a desk frame, which caught desks described as having "a steel frame", and a monitor riser was counted as a desk converter. We read all 50 rows and fixed the rules. A short manual check of what code classified is cheap insurance.

The same flow in Python, with an API key in LOVELY_API_KEY. run_tool starts a tool and waits for its items; ask sends one chat request on the same key.

import json, os, requests

BASE = "https://lovelyrouter.com/v1"
H = {"Authorization": f"Bearer {os.environ['LOVELY_API_KEY']}"}

def run_tool(tool, input, max_results):
    r = requests.post(f"{BASE}/tools/{tool}/run", headers=H, json={"input": input, "max_results": max_results, "wait": 120}).json()
    while r["status"] == "running":
        r = requests.get(f"{BASE}/tools/runs/{r['id']}?wait=120", headers=H).json()
    return r.get("items", [])

def ask(system, data, model="claude-sonnet-5.5"):
    r = requests.post(f"{BASE}/chat/completions", headers=H, json={"model": model, "max_tokens": 12000, "messages": [
        {"role": "system", "content": system}, {"role": "user", "content": json.dumps(data)}]}).json()
    return r["choices"][0]["message"]["content"]

best = run_tool("amazon-bestsellers", {"urls": ["https://www.amazon.com/Best-Sellers-Office-Products-Office-Desks-Workstations/zgbs/office-products/12900751"]}, 50)
reviews = run_tool("amazon-reviews", {"input": [{"asin": p["asin"], "domainCode": "com", "sortBy": "recent", "maxPages": 2} for p in best[1:6]]}, 100)
posts = run_tool("reddit-search", {"queries": ["standing desk worth it", "standing desk wobble problem"], "maxPosts": 25, "timeframe": "year"}, 50)
trend = run_tool("google-trends", {"mode": "keyword", "keyword": "standing desk", "predefinedTimeframe": "today 5-y", "geo": "US"}, 1)

tags = ask("Tag each review with up to 3 themes from this list: ... Return a JSON array: {id, themes, quote}. Copy quotes exactly.",
           [{"id": i, "stars": r["rating"], "text": r["text"]} for i, r in enumerate(reviews)])
# Count themes in code, check each quote is in its review, then:
# brief = ask("Write a market brief. Use only these numbers and cite each one.", {...counts, prices, trend by month...})

Swap the category URL, the review products and the search phrases and the same script researches any product category. Each run of this size costs well under a dollar. To let an AI agent run it end to end instead, connect it to the MCP server: it finds and runs the same tools itself.