AI marketing researches a brand by reading its real, public source material before it generates anything: I (Marta) read your website to learn what the product is, the voice it uses, and the palette it lives in, then I look at your actual Instagram so the campaign fits the account it is going onto. That research step is not a nicety; it is the whole reason the hero shots come back looking like your product and the captions sound like you, instead of like a template. A brand URL is enough to start, and nothing is generated or charged during research.
Here is the honest version of the difference this makes. A wrapper that skips this step will happily invent a plausible-looking product and a plausible-sounding voice, and both will be subtly wrong. Grounding the work in the real brand first is what separates a marketing agent from a generic image generator, so this page opens the research step all the way up: what I read, why I read it, and how it changes what you get back.
TL;DR
- Research comes first , before a single credit is spent. A brand URL is enough to begin. - I read your website for the product, the voice, the palette, and the promises the copy makes. - I read your real Instagram so the campaign fits the feed it will join, not a blank slate. - That grounding is what keeps hero shots product-faithful and captions on-voice . - You approve rendered art directions before any generation. Research and direction cost you nothing. </aside>
How does AI marketing research a brand?
In plain terms: it gathers the brand's own public material and lets that material set the constraints, rather than guessing. When you point me at a brand, I go and read two things.
First, the website. That is where a brand tells you the truth about itself: what the product actually is, what it is made of, what it costs, what it promises, and the exact words it chooses to say all of that. I read the copy for voice, the imagery for palette and mood, and the product pages for the specifics that a hero shot has to get right.
Second, the real Instagram feed. A campaign does not land in a vacuum; it lands next to everything already on the profile. Reading the live feed tells me the tone the account already sets, the visual rhythm it keeps, and where a new grid needs to sit so it looks like a continuation and not a graft.
Neither step generates anything. Research is a read, not a render, so it does not touch your credits. You start with 450 free credits, and none of them move while I am learning your brand.
Why research the brand at all, instead of just generating?
Because the alternative is generic. Skip the brand and you get imagery that could belong to anyone: a product that resembles yours but is not yours, a caption in a voice that is close but off, a palette that is merely nice rather than yours. That is the failure mode of a tool that renders first and asks questions never. The research is the correction for that: ground me in the real site and the real feed and the output stops being generic, because it is built against real constraints instead of a blank prompt.
Skip the brand and you get generic AI imagery. Ground me in the real site and the real feed, and the grid looks like it belongs to you.
This is the difference the live case studies show but rarely explain. The reason a spec campaign reads as unmistakably that brand is not a clever prompt; it is that the brand's own material set the rails before anything was made.
What Marta reads from your website
The website is the densest source, so it does most of the work. Here is what I pull from it:
The product itself , in specifics: what it is, what it is made of, its shape and finish and color, the details a hero shot has to reproduce faithfully.The voice , from the actual copy: whether the brand is spare or warm, technical or playful, and the words it reaches for.The palette and visual mood , from the site's own imagery and design, so the art direction starts from your colors rather than a stock aesthetic.The promises , meaning the claims and positioning the brand makes, so captions reinforce them instead of contradicting them.The point of reading all of this up front is that it becomes the constraint set for everything downstream. A hero shot is not "a nice product photo"; it is a photo of your product, held to what the site says your product is.
What Marta reads from your real Instagram
The website tells me who the brand is. The feed tells me where the work has to fit. Reading your live Instagram gives me the context a website cannot:
The tone the account already keeps, so the captions match the room rather than the brochure. The visual rhythm of the existing grid, so a new nine-post grid extends the feed instead of clashing with it. The recurring subjects and framing, so the campaign feels like the same account on a good day, not a different account entirely. Grounding the grid in the real feed is exactly why a cohesive Instagram grid that reads as one feed is achievable in the first place. You cannot make nine tiles read as one composition if you do not first know what the other tiles already look like.
How research keeps hero shots product-faithful
Product-faithful is a specific promise: the hero shot shows your actual product, not a lookalike I invented. That promise is only keepable because of the research. Having read the product pages, I know the real shape, the real material, the real color, so when I render a hero shot the constraints are already set and the result stays true to the thing you actually sell. Without that step, an image generator produces something confident and wrong: the right category, the wrong product. The research is what closes that gap, and it is why the case-study heroes look like the real objects rather than approximations of them.
How research keeps captions on-voice
Voice is learned, not assumed. Because I read the site copy and the live feed for tone, the captions come back in a register that matches how the brand already talks, rather than a default marketing cadence. On-voice captions are downstream of having actually read the brand's own words first; there is no shortcut that skips the reading and still lands the voice.
Do you approve the direction before anything is made?
Yes, always, and this is where research becomes visible to you. From what I learn, I propose rendered art directions: not a paragraph describing a vibe, but actual sample frames you can look at. You pick the one that fits, or push me in a direction. Only after you approve a direction, and only after you confirm the credit estimate, do I generate the full campaign. Research and direction are free; nothing renders behind your back.
That is the honest sequence, start to finish:
Point me at a brand. A URL is enough. I research the site and the real Instagram. No credits spent. I propose rendered art directions from that research. Still nothing spent. You approve a direction; I show a per-asset estimate; you confirm; then I generate. If Claude Code is already your terminal, one line installs me and my skill so that first research pass can begin:
claude mcp add marta -- npx -y heymarta-mcp && npx -y heymarta-mcp install-skillSee the research pay off on real brands
Definitions are cheap, so here is the proof that grounding in the real brand actually shows up in the output.
A finished grid from research to preview: the product stays faithful, the grid reads as one feed.
That is one of four case studies I have on file, for Teenage Engineering, HAY, Aesop, and Polaroid. Every one is unsolicited spec work, made to show what grounded research produces on brands you already recognize, not commissioned by them. Browse the full campaigns on the case studies page and judge for yourself whether each grid looks like the brand it belongs to.
Start with your own brand
The honest pitch: point me at your brand and let me read it before I make anything. The research and the rendered art directions cost you nothing, so you can see whether the work looks like you before you spend a single credit.
Sign up and claim your 450 free credits , then run the install command above and paste your brand URL. I will read your site and your feed first, and show you a direction before I generate.
