What Actually Happened When We Used AI to Generate 1,000 Product Descriptions: Lessons from a Real Deployment
When Grosvenor Flooring came to us, the brief was clear: their product catalogue had grown faster than their ability to create content for it. Hundreds of products had no descriptions. Others had one-line placeholders copied from supplier data sheets, the kind of content that does nothing for search visibility and actively undermines trust when a customer lands on the page. The team was small, there was no budget to hire a content operation, and the backlog grew every time a new product range arrived.
This is not a pilot story. This was a live WooCommerce storefront, real customers, real search rankings, real revenue at stake. I want to write honestly about what the deployment actually looked like, because most AI case studies skip the rough edges and lead with the headline number. This one will not.
What we actually built (it is more than writing descriptions)
The first thing I need to address is the framing problem. When retailers ask about AI product content, they almost always mean writing descriptions. What we built at Grosvenor Flooring was closer to a content operations pipeline, where writing is one output among several.
The full implementation covered six capabilities:
Product descriptions generated from attributes, images, and category context. Each description followed a defined schema (key attributes in structured order, SEO-relevant language, no generic filler) applied consistently across the entire catalogue.
Lifestyle imagery using generative AI to produce room scene visualisations. Grosvenor Flooring had no photography studio. Generative imagery meant every product could have an aspirational context image, something previously reserved for the category’s largest operators.
AI Floor Finder, a visual similarity search tool. A customer uploads a photo of a floor they have seen somewhere (a showroom, a friend’s house, a magazine) and the AI finds the closest matches in the catalogue by comparing images. No product name, category, or specification needed. The customer shows the system what they want, and it finds it.
AI Room Visualiser, a generative staging tool that lets customers see a specific flooring product in their own space before purchasing.
Mood Board Analyser, an LLM-powered tool for interpreting design intent from customer images and free text, then routing that intent to relevant products.
Omnichannel content sync to ensure content generated for the web flowed consistently to Google Merchant, email, and WhatsApp.
What worked from day one
Speed was the immediate win. The rate at which the platform processed products made it clear the backlog would be cleared in hours, not the months a manual process would have required. For a small team staring at a growing list of products needing attention, that timeline visibility mattered.
Consistency was the less obvious but arguably more important win. Manually assembled catalogues accumulate stylistic entropy: some descriptions are thorough, some are thin; some use technical terminology, others avoid it. The schema-driven approach meant every product received the same structure, the same attribute depth, the same content standard. That consistency matters to search engines, which reward structured content, and to customers, who build trust through predictable product pages.
Attribute completeness improved substantially. Products with missing specification fields were populated from image analysis and category inference. A plank that previously had no wear rating, installation method, or underfloor heating compatibility recorded now had all three, extracted automatically rather than filled in manually.
What needed iteration: the honest account
Brand voice calibration took around two to three rounds of refinement before the output consistently matched Grosvenor Flooring’s tone.
The challenge was not technical. The model produced grammatically correct, factually accurate descriptions from the first run. The challenge was editorial: the initial output was accurate but slightly generic. The knowledge base needed more precise seeding with the brand’s specific vocabulary, the terms they use for product qualities, and the phrases they actively avoid. It read like a competent contractor who understood the brief but had not yet absorbed the brand.
This is not a failure mode unique to AI. It is the same onboarding gap you encounter with any new copywriter. The difference is that once the knowledge base is correctly configured, every subsequent product benefits from that configuration immediately. The iteration cost is front-loaded, not recurring.
Edge cases required human review. Very large format tiles, discontinued product ranges, and products where supplier photography was low-resolution all needed a second pass. The platform’s ability to extract attributes from images is bounded by image quality. We flagged these for manual handling rather than pushing them through the pipeline at lower confidence, which added time on those specific products but protected the overall output quality.
What surprised us
Taxonomy classification accuracy. We expected to spend significant time correcting category assignments, particularly for products that could sit in multiple categories. The accuracy was materially better than anticipated, which accelerated the omnichannel sync, a step that depends on clean taxonomy to route products correctly to Google Merchant categories.
The AI Room Visualiser’s commercial impact. We built it as a product feature. What we did not predict was how substantially it changed the shape of the customer journey. Customers who engaged with the Room Visualiser converted at a higher rate. Flooring is not an impulse purchase. The ability to visualise the product in a real space reduced the uncertainty that was previously causing hesitation, in a way that even good photography could not fully replicate.
The results: why the 976% figure is meaningful
Grosvenor Flooring achieved 976% online revenue growth over the period of the merchi.ai deployment. That figure appears in our National AI Awards 2026 Finalist entry and the Grosvenor Flooring case study.
To be precise about what it represents: it is the measured increase in online revenue attributable to the channel improvements the platform enabled: improved search visibility from structured product content, better product discovery from the AI Floor Finder, and higher conversion from the AI Room Visualiser and lifestyle imagery.
No single feature produced that number. It is the product of a compounding pipeline. Remove any component and the figure changes. That is why we describe this as a systems result rather than a conversion rate story. The full ROI breakdown breaks down the four ROI drivers and the causal mechanism in detail.
What we would do differently
Two things.
First: brief the knowledge base more thoroughly before running the first batch. Two to three rounds of voice calibration is manageable but avoidable. A thorough upfront brief covering tone, vocabulary, sentence structure, and terms to avoid would have compressed that to one round. The knowledge base is the AI’s editorial context. Treating it like a configuration file rather than a content brief is a mistake.
Second: define the schema before starting, not after. The content schema evolved during the first phase of the deployment as we learned more about how content would be used across channels. That evolution created rework on early batches. For every new deployment we run now, the schema is defined in full before generating a single product. It sounds obvious. Most content projects make the same mistake.
If your catalogue has a content backlog, products not performing in search, or a team that cannot keep up with new product velocity, the approach we used at Grosvenor Flooring is repeatable. See product content at scale for how the pipeline handles different catalogue types. Book a 20-minute walkthrough at our scheduling page. Or start a 30-day free trial to run the platform on your own catalogue.
Frequently asked questions
Does AI product content actually work?
Yes, in production at a named retailer. Grosvenor Flooring achieved 976% online revenue growth over the merchi.ai deployment period. The growth came from improved organic search visibility from structured product descriptions, better product discovery from the AI Floor Finder, and higher conversion from the AI Room Visualiser and lifestyle imagery. These results are documented in the Grosvenor Flooring case study and recognised by the judges of the National AI Awards 2026, for which merchi.ai was named a finalist in the AI SME Business of the Year category.
How long does it take to get AI product content right?
Initial processing is fast: a catalogue of 1,000 products can be processed in hours rather than weeks. Brand voice calibration typically takes two to three iteration rounds, adding days rather than weeks. The longest part of a deployment is upfront: knowledge base briefing, schema definition, and establishing the brand’s vocabulary and tone preferences. Investing time in that stage compresses the iteration cycle significantly. Edge cases (unusual products, low-quality supplier images) should be flagged for separate manual review rather than held up in the main pipeline.
What kinds of products does AI product content work best for?
Products with well-structured attribute data and reasonable image quality are the easiest starting point. The pipeline extracts attributes from both data sheets and images, so richer source material produces richer output. The Grosvenor Flooring deployment covered flooring across multiple categories (LVT, laminate, and wood) which required taxonomy classification as well as description generation. Taxonomy accuracy was one of the surprises: classification across category boundaries was more reliable than expected.
What happens with products that have unusual or complex attributes?
Products outside the standard attribute set are flagged for human review rather than processed at lower confidence. In the Grosvenor Flooring deployment, this covered very large format tiles, discontinued ranges with incomplete supplier data, and products where image quality limited attribute extraction. The pipeline handles clear cases automatically and surfaces edge cases for manual handling. Partial automation with human review on the exceptions is more accurate than full automation across all cases.
How does merchi.ai differ from a general-purpose AI writing tool?
General-purpose AI writing tools produce text from a prompt. merchi.ai produces structured product content within a defined schema, using product attributes and images as inputs. The output is consistent across every product, schema-compliant, and tagged with the AI Provenance Protocol for machine-readable provenance tracking. The platform also integrates taxonomy classification, lifestyle imagery generation, and search tooling. The difference is the difference between a word processor and a content operations platform.