AI Content Optimiser

AI Content Optimiser puts a ✨ Optimise (and ✨ AI Write) button next to the fields you actually write — product name, short & long description, meta title/description, category descriptions, and CMS content. Click it and Google Gemini drafts SEO-grade copy in place, using prompts you can read and edit, with strict factual-accuracy guardrails and server-side HTML sanitisation.

Compatibility

Magento

Open Source 2.4.9 GA (and later 2.4.x).

PHP

Tested on 8.4 and 8.5.

AI engine

Google Gemini (your own API key), with automatic fallback model.

Scope

Per-store enable + locale, so each store view writes in its own language.

What it adds

The ✨ buttons appear directly on the standard Magento edit forms — no new screen to learn:

Product

• Name — ✨ Optimise (≤ 80 chars, keyword front-loaded)
• Short Description — ✨ Optimise (~36-word standalone summary)
• Description — ✨ AI Write (400–800 words HTML: intro, H2s, list, table)
• Meta Title — ✨ Optimise (≤ 60 chars)
• Meta Description — ✨ Optimise (≤ 155 chars)

Category

• Description — ✨ AI Write (200–400 words HTML, above-the-fold category intro)
• Meta Title / Meta Description — ✨ Optimise

CMS page

• Content — ✨ AI Write (structured HTML, headings/lists/tables)
• Meta Title / Meta Description — ✨ Optimise

Give it context

The AI Write fields have a small context box — paste specs, a competitor paragraph, or a few bullet points and the model uses them as source facts. The more real detail you give it, the better (and more accurate) the result.

Quick start

1. Get a Google AI key from Google AI Studio (aistudio.google.com/apikey).

2. Stores → Configuration → AgenticEcom · AI → AI Content Optimiser → AI Engine: set Enable = Yes and paste your Google AI API Key (stored encrypted).

3. (Optional) Pick a Primary Text Model and a Fallback Text Model (used automatically if the primary returns a 503/429).

4. Open any product, category or CMS page and click a ✨ button.

You control the prompts

Every button’s behaviour is a prompt you can read and edit per field — no black box.

Under Product / Category / CMS Field Prompts, each field has its own instruction textarea (character targets, structure, keyword rules). Tune them to your brand voice once and every generation follows your rules. Prompts are store-scoped, so different store views can have different instructions — and different languages.

Built-in writing standard

On top of your per-field rules, every generation is wrapped in a strict house style: E-E-A-T and BERT/GEO optimisation, plain active-voice English, and a long banned-phrase list (no “game-changer”, “cutting-edge”, “seamless”, “elevate”…) so the output never reads like generic AI filler.

Accuracy & safety

This is the part most “AI description” tools get wrong. The optimiser is built so you can trust what it writes onto your live catalogue.

No invented facts

The prompt forbids guessing or inventing specs, dimensions, materials, numbers, prices, certifications or reviews. If a detail isn’t given or genuinely known, it’s omitted — saying less beats saying something wrong.

HTML is sanitised

Rich-text output (descriptions, CMS content) is cleaned server-side before it’s returned: an allow-list strips <script>, event handlers, javascript: URLs, SVG and other stored-XSS vectors, while keeping headings, lists, tables, links and images.

Your key, encrypted

The API key is stored encrypted and sent to Google in a request header, never in a URL — so it can’t leak into proxy or access logs.

Permissioned

Generation sits behind its own AgenticEcom_AiContent::generate ACL resource and the config behind ::config, so you decide which roles can use it.

FAQ

Does it overwrite my field automatically?

It drafts into the field so you can review and tweak before saving — nothing is written to the catalogue until you click Save. Treat it as a fast first draft from a specialist copywriter.

Will it make up specifications?

No — that’s the point of the accuracy guardrails. The model is instructed to use only facts you provide in the context box or that are genuinely established about that exact item, and to omit anything it isn’t sure of rather than invent it.

Can different store views write in different languages?

Yes. Enablement, prompts and locale are store-scoped — edit a product in a German store view and the generator responds in German, following that store’s prompt overrides.

What happens if Gemini is busy?

If the primary model returns a 503/429, the client automatically retries with your configured Fallback Text Model, so a transient outage doesn’t block your work.

Is the generated HTML safe to publish?

Yes — all rich-text output is run through a server-side sanitiser that allow-lists safe formatting tags and strips scripts, inline event handlers, unsafe URLs and SVG before the HTML ever reaches your storefront.