NotebookLM Data Tables: I Fed It a Messy Spreadsheet
NotebookLM Data Tables promises to turn messy data into clean tables. I tried it with a real messy spreadsheet. The catch is at the input, not the AI.
Create
July 7, 2026 · 5 min read
AI product photos are cheap enough that skipping the studio feels obvious: Pebblely's free tier gives you 40 images a month, and Photoroom's Pro plan starts at $7.99. What nobody selling those subscriptions shows you is which products actually come out looking sellable, because every tool's own gallery only shows the win. So instead of generating one flattering batch myself and calling it a test, I spent a day pulling together every independent before/after I could find: strangers testing the same tools on perfume bottles, diamond rings, coffee mugs and sneakers, using their own products. Version one, according to nearly every one of them, was terrible. Here's the honest map of what survives and what breaks, and why the failures cluster around exactly the products you'd expect to matter most.
Yes, for roughly half a typical catalog, and the line falls exactly where you'd guess: simple, opaque, text-light products cross it easily, and anything reflective, transparent or covered in a logo usually doesn't. That split holds up across both independent testers and the tool companies' own comparison posts, which is rare agreement for a space this full of marketing. The products that struggle share one trait: they need the AI to get a specific physical detail exactly right, whether that's a diamond's facet count or the second letter of a brand name, and generative models are still bad at exactness. Simple geometry has nothing exact to get wrong.
Three tools cover most of what a small store needs, and none require a paid plan to start. Photoroom's free tier includes 250 exports a month; Pro is $7.99 a month for AI-generated backgrounds and batch export, and Max jumps to $26.99 for better models and Shopify integration. Pebblely's free tier gives you 40 images a month, with paid tiers from $19 to $39. Claid.ai starts with a 50-credit trial, then $9 a month for unlimited enhancements or $39 for custom templates and outpainting. Budget an afternoon for a real test batch, not a full day: you're checking one thing, whether your specific products survive.
Plain, geometrically simple products won almost every independent test I found. One blogger who ran more than 20 different AI photo tools against her own products found the pattern held batch after batch: a water bottle and a pair of sneakers came out looking shot in a real studio, with clean backgrounds and shadows matching the product's actual shape. Boxed goods, plain clothing, and basic electronics show up the same way across other comparisons. The common thread isn't the category. It's the absence of anything the AI has to render exactly: no facets, no fine print, no complicated reflections.
Glass and polished metal are where the same tools stop agreeing with reality. Orbitvu ran 5 different AI models against 4 perfume bottle photos and found every model changed something a customer would notice: bottle proportions shifted slimmer or bulkier than the source photo, reflections looked studio-perfect instead of matching the actual scene, and in one output a logo's ampersand came out as a dollar sign. Jewelry fares worse. Masonry AI tested one diamond ring across 4 models and found a plain solitaire held its shape, but complex settings, pavé bands, the exact number of prongs, and hallmark inscriptions all changed between generations. A separate comparison by FormaNova on a similar ring showed how far that drift can go: one tool flattened the design into what the reviewer called a thin plain gold band, another produced something closer to a rectangular brick than a ring.
Nobody sets out to sell a ring shaped like a brick, and yet here we are.
Here's the weird part: it's not just jewelry and glass. Pebblely's own team has written about why AI-generated text keeps coming out wrong, and the reason is uncomfortable for anyone selling the tool: most AI image generators don't actually read text at all, they draw shapes that resemble letters, which is why a label can look right at a glance and say something subtly wrong up close. The same 20-plus-tools blogger found other tells piling up outside the logo problem: a hand holding a book with four fingers instead of five, steam rising off a coffee mug like a smoke machine had gone off next to it, and a hat floating a visible gap above where a head should sit. Even academic research on text-to-image models agrees that hands remain a documented, unsolved failure mode, which is one more reason to keep people's hands out of your product shots entirely rather than trust a prompt fix.
The pattern across every source, vendor-run comparisons included, is consistent enough to act on: plain, opaque, text-free products are close to a solved problem, and reflective, transparent, or text-heavy products are not, no matter which tool you pick. Test a category with four checks: upload one plain product and one reflective or textured product to a tool's free tier, generate three variations of each, zoom to 200% on every letter of any text or logo and read it out loud, then compare the reflection angle to the actual light in your product shots. That costs an afternoon, and it tells you more than any tool's marketing page will.
| Tool | Free tier | Paid tier (from) | Best for |
|---|---|---|---|
| Photoroom | 250 exports/month | $7.99/month | Batch backgrounds, Shopify sync |
| Pebblely | 40 images/month | $19/month | Bulk lifestyle scenes |
| Claid.ai | 50-credit trial | $9/month | Enhancement, templates |
If you're rebuilding the product page around new photos, the copy is the other half of that job. I've written separately about how to write blog posts with AI, and tested which AI writing tools actually hold up for shorter product descriptions.
Skip AI product photos if most of your catalog is jewelry, glassware, mirrored surfaces, or anything with fine print that has to be legally accurate, like ingredient lists or warranty text. The failure mode isn't a slightly off photo. It's a photo that looks fine at thumbnail size and wrong the moment a customer zooms in, and for regulated text that's a liability, not a style problem. A real photographer, or at minimum a careful manual retouch, still earns their fee here.
Not for a full batch, no. For anything simple, plain, and free of text, the odds are good enough that the studio question barely feels like a question anymore. For everything else, glass, metal, jewelry, logos, I'd rather pay a human once than explain to a customer why their ring looks like a brick. The tools are good. The trust has to be earned product by product, not granted to the category.
NotebookLM Data Tables promises to turn messy data into clean tables. I tried it with a real messy spreadsheet. The catch is at the input, not the AI.
I gave five AI writing tools one brief: write my bio. Each broke differently. Where ChatGPT, Jasper, Copy.ai, Grammarly and Sudowrite fall short.
A Sora 2 test review that skips the hype: a verified map of where its physics promise breaks (hands, flips, text) and why the app closed in 2026.