What the AI bases content generation on

Fexa AI does not generate "generic" content. Before writing a single line, it reads a set of precise data from your store and Google Search Console. Here is exactly what is injected into each generation.

Prerequisite: Make sure you have configured your store and launched a full SEO audit before generating content.

The 8 data sources injected

1. Complete product sheet (PrestaShop)

Via the MCP Bridge, the AI reads in real time:

• Product name, brand, parent category
• Tax-exclusive price
• Existing short and long description (to improve rather than start from scratch)
• Product reference

Technical specifications (PrestaShop Features)

Fexa AI reads the Features table of the PrestaShop back-office (Catalog > Products > Features tab). Specifically:

• If your features are filled in (e.g., Weight: 2.5 kg, Material: Aluminum, Warranty: 2 years), the AI integrates them verbatim in a <h3>Technical specifications</h3> section of the long description.
• If the table is empty or not filled in, the AI receives an explicit warning and does not invent any technical data. It focuses only on the name, category, and existing description.

Absolute rule: Fexa AI cannot write what is not in your back-office. The more complete your product sheets are, the more precise and differentiating the generated content will be.

PrestaShop Combinations / Variations

Fexa AI reads the product's combinations (Catalog > Products > Combinations tab):

• If combinations exist (e.g., Color: Red, Blue, Black | Size: S, M, L, XL), the AI mentions them naturally in the description (e.g., "Available in Red, Blue, and Black, this product adapts to all styles").
• If no combinations are configured, no variants are mentioned — never invented.

Combinations are also used in the meta_description to signal the richness of the offer (e.g., "Available in several colors — Fast shipping").

2. Google Search Console (GSC) Data

If you have connected your GSC, the AI receives for each product:

• Top 10 keywords bringing traffic to this URL
• Current position on Google for each keyword
• Number of clicks and impressions
• CTR (click-through rate)

This data allows the AI to:

• Naturally integrate keywords that are already working
• Detect Striking Distance keywords (position 10-20) and prioritize them in H2/H3 headings
• Rewrite an unattractive meta_title if the CTR is low despite a good position

3. SEO Audit Flags

The SEO score of each article is accompanied by warning flags (seoAuditFlags) calculated during the scan:

• MISSING_META_TITLE — missing title
• SHORT_DESCRIPTION — description too short
• DUPLICATE_CONTENT — content similar to another article
• LOW_KEYWORD_DENSITY — insufficient keywords
• MISSING_ALT — images without alternative text

The AI receives this list and specifically addresses each detected issue.

4. Cannibalization Signals

If two articles in your store target the same keyword, the AI receives:

• The [CANNIBALIZED] flag on the keyword concerned
• The indication of who is the authority article ([AUTHORITY]) vs. the internal competitor ([COMPETITOR])
• The name of the conflicting article

It then differentiates the content or adds an internal link to the authority page.

5. Existing Images

For each product image, the AI receives:

• The image identifier (id)
• The current caption (if filled in)
• The image URL

It generates an optimized ALT tag per image (max 100 characters, containing product keywords), returned in the image_alts field.

6. Available Internal Links

To build the semantic mesh, the AI receives a list of:

• 20 sibling products from your catalog (name + URL)
• Categories (used only if no products are available)

It selects 1 to 3 relevant products and integrates them naturally into the text, never as a list at the end.

7. Editorial Tone and Brand Identity

In the shop settings (Settings > AI Preferences page), you define:

• Tone: Expert, Luxury, Friendly, Commercial, Neutral
• Identity: a free-form text describing your brand (e.g., "We are a family-owned SME specialized in urban bicycles since 1998")

These two elements are injected at the top of the system prompt so that each text sounds like your voice, not a robot's.

8. Target Language

The ISO code of the article's language (fr, en, de, etc.) is translated into a full name and injected with the absolute instruction: "YOU MUST GENERATE ALL TEXTS IN [LANGUAGE]". There is no possible ambiguity about the output language.

The generation engine and the anti-template memory

Which model writes your content?

By default, Fexa AI relies on Gemini 3.1 Pro, Google's premium engine, chosen for the quality and naturalness of its writing. If this model is momentarily unavailable, the platform automatically switches to a fallback model rather than failing — your generation succeeds in every case, and the error message (if any) stays honest and explicit.

The signature memory: ending the "template effect"

When generating hundreds of product sheets, the risk is that the AI keeps reusing the same turns of phrase ("Discover our…", "Combining elegance and performance…"): the infamous template effect that betrays mass-produced content, both to visitors and to Google.

To avoid this, Fexa AI keeps a signature memory at the store level. With each generation, it records:

• the opening sentences already used;
• the recurring expressions (word groups);
• the overused descriptors (stylistic keywords), in all their forms.

These elements are then banned or made scarce in subsequent generations, and the AI rotates the editorial angle from one sheet to the next (customer benefit, usage context, craftsmanship, comparison…). The result: even on a large catalog, each sheet keeps its own voice, with no mechanical repetition.

Note: this memory is agnostic — it works whatever your industry or language, and it sharpens over the course of your generations.

Per-language anti-clichés & editor pass

Beyond the signature memory, Fexa AI bans the clichés specific to each language — not just the English tics. In English, worn-out turns of phrase like "elevate your…", "a true…", "at the heart of…" or "sensory journey" are explicitly avoided (the same goes for French, Spanish, and Italian), and the abstract feeling is replaced by a concrete fact drawn from your data (a rating, a material, a real measurement rather than an "undeniable presence").

Finally, if a first draft still turns out too clichéd or too monotonous (uniform sentence rhythm), an "editor" pass triggers automatically: a second pass rewrites the text to vary the rhythm and remove the tics — without ever touching your specifications, figures, or links — and only if the result is measurably cleaner (otherwise the first version is kept). This pass only triggers on the minority of sheets that need it and does not consume any extra credits.

What the AI Does Not Do

• It does not invent technical features: if the Features field is empty in PrestaShop, no specs will appear in the generated text.
• It does not invent combinations: if no combination is configured, it will not write "available in 3 colors".
• It does not copy content from other sites or supplier sheets.
• It does not generate in one language and then translate — each language is a native, independent generation.
• It does not fabricate prices, delivery times, or warranties not present in the data.

When your product lacks data (the "data-poor" strategy)

What happens with a product that has no features, no combinations, and an empty or very short description? Fexa AI does not paper over the gap with generic filler, and it never invents specs. Instead, it switches to an honest strategy:

• It anchors only on what is reliable: the name, the brand, the category, and the real search intent (your Google Search Console keywords).
• It writes what is useful: what the product is for, who it's for, where it sits in its category, and a FAQ built from the real buying questions.
• It favors short and precise over long and vague — no hollow paragraphs just to "add volume".

Note: for a product that is nothing but a name (zero data), even the best AI can only write at the "category / intent" level. The real quality lever remains enriching your PrestaShop catalog.

The "To enrich" card on your dashboard

To spot these products without hunting them down one by one, your dashboard shows a "To enrich" card: it counts the sheets whose data is too thin for a rich description (thin content, or no features and no combinations) and sends you straight to the product to complete. Fill in its features/combinations in PrestaShop, then regenerate — quality climbs immediately.

And when a sheet is generated in "data-poor" mode anyway, a warning appears on the suggestion to flag it to you in full transparency.

Improving Output Quality

Practical tip: For a catalog of 500 products, start by filling in the features of the top 50 best sellers, then generate in batches. You will see the quality difference immediately.

Going Further

• Generate SEO-optimized product descriptions
• Create smart internal links
• Generate title and meta tags
• Understand the credit system
• Track your scores' evolution