Gamma with Claude and Linkeddit: measured competitor data in, finished decks out

Gamma is the designer. Its MCP server, hosted at https://mcp.gamma.app/mcp, creates presentations, documents, webpages and social posts, reads and searches existing ones, exports to PDF, PPTX or PNG, and browses themes and folders. It knows nothing about what your competitors shipped this week or what ChatGPT says about your category.

Linkeddit is the source of facts. Compete tracks competitor moves and their customers' reviews. Answer Radar asks the five engines your buying questions and records who is named, recommended and cited. Both come with one plan at $60 a month.

Claude sits between them. It decides what belongs on a slide, and Linkeddit writes no copy or design of its own, so every figure can be traced to a measured run.

Gamma's docs list generating from scratch, generating a multi-page gamma, generating from a template, editing an existing gamma, exporting, reading and searching gammas, and browsing themes and folders. Generation is asynchronous: Claude starts it, then polls get_generation_status until it is completed or failed.

Gamma says the connector is available on all plans, that generations and edits spend credits, and that reading, searching and exporting do not. It also says edit_gamma only appears when editing is enabled for the account, so check the tool list before relying on it. The older directory listing we checked shows only three tools, so trust Gamma's own reference.

People do try to get Claude to produce visual output and struggle. A Canva user on Reddit wrote: "I’ve asked Claude several different ways how to do this, and it gives me the same result as one of the above, which does not work well." Gamma's own connector gives Claude a direct route to a deck, and Linkeddit supplies the data it puts in it.

The first is an AI visibility report: pull per engine presence, recommendation rate, share of voice and cited sources from Answer Radar, then have Claude lay them out in a Gamma deck. It is covered on this page, including how to label an engine that returned no usable data.

The second is a competitor analysis presentation built from the weekly Compete brief, with a source link on every claim. The third is a set of sales battlecards, one per rival, built from the brief and the customer reviews Linkeddit has stored.

Read every figure against the source. Ask Claude to read the finished deck back with read_gamma and compare it with the Linkeddit output. A layout step can drop a number or shorten a quote. Gamma documents textMode preserve for text you want kept exactly as written.

Say what the numbers are. Linkeddit measures through official provider APIs with grounded web search and states in each result that this is a controlled proxy, not an exact copy of what a person sees in consumer ChatGPT or Google. Put that line in the footer of any report that leaves your team.

Gamma says its MCP server is available on all plans, and that generating and editing spend credits in your Gamma workspace while reading, searching and exporting do not. Those credits are Gamma's, so check your plan there.

Linkeddit is one plan at $60 a month. It covers Compete, Answer Radar and Search, and it is the data side of every guide here.

Profound's Gamma integration doc shows the shape. Profound sends a structured prompt to Gamma, for example a 10 slide executive summary with a title slide, a summary slide, data breakdown slides and a recommendations slide, and fills it with a visibility score, share of voice and top topics. Its doc names the use directly: client-ready visibility reports and executive decks built from answer engine performance data.

The numbers must be measured per engine and carry their sample size. An average across engines hides the case where ChatGPT names you in 16 of 18 answers while Claude names you in 5 of 10. Put the numerator and denominator on the slide.

The reporting problem is common. An SEO consultant on Reddit asked: "I’m curious how other SEOs are tracking AI visibility without turning it into a spreadsheet of manually checked prompts." A deck built from stored run results answers that with a repeatable report instead of a hand-checked list.

On Linkeddit's test project, plausible.io, the latest results had 7 engine segments, of which 5 were usable. ChatGPT named the brand in 16 of 18 admissible answers and recommended it in 15. Plausible held 20.5% share of voice there, ahead of Fathom at 16.7% and Matomo at 15.4%. Google AI Overviews named it in 15 of 15 answers, but showed an overview in only 16 of 21 samples, so 5 were absent.

Perplexity returned no admissible samples: all 21 were errors. A deck that quietly shows 0% for Perplexity says something false. The correct slide says not measured this run and gives the reason.

Ask Claude to call answer_projects_get, then answer_results_get, and answer_runs_get for the latest run dates. For each engine it lists presence_rate, recommendation_rate, the top five vendors by share_of_voice and the top cited domains, and puts any engine with usable false on a slide called Not measured this run, with the reason given.

Have Claude write a 10 slide outline for leadership and wait for your approval before any design exists: a summary, one slide per measured engine, who is recommended instead of you, which sources the engines cite, and three next steps. Check every number against the Linkeddit output.

After you approve, Claude calls generate in Gamma with format presentation, the outline as inputText and textMode preserve, which Gamma documents for text you want kept as written, then polls get_generation_status. The generate tool also takes a card count, an optional theme and an exportAs value of pptx, pdf or png. Open the deck, compare each figure with the source, then ask for export_gamma as PDF. Gamma's guide to turning reports into presentations suggests an executive summary first, key findings with visual weight, methodology kept short and recommendations as clear actions.

To refresh, pull answer_results_get again, list what moved against the previous run and change only the figures. Reading stored results is a read of data already measured, so a refresh does not need a new measurement run.