Right here’s how the small-but-mighty Ahrefs product advertising workforce makes use of AI to automate their work.
Andrei leads product advertising at Ahrefs, with a small workforce protecting copy, webinars, partnerships, and paid promotion for all of Ahrefs (together with the handfuls of updates we ship every month).
Constance works on the product workforce, operating all of our webinars, writing our assist middle documentation, and rather more in addition to (try Constance’s webinars on YouTube).
Andrei and Constance constructed a product advertising workspace in Agent A that automates their workflows and builds customized purposes. Listed here are the eight instruments they lean on the toughest as PMMs.
What’s Agent A?
Agent A is a advertising agent from Ahrefs—an AI assistant with direct entry to the total Ahrefs dataset that may perform advertising duties autonomously, fairly than simply reply questions.
Agent A consists of:
- Unrestricted entry to Ahrefs endpoints. Each endpoint we use to construct Ahrefs is out there, together with many you can not attain by way of API or MCP.
- Critical tech stack beneath. Postgres for state, Flask for UIs, an OpenRouter proxy with 300+ fashions, internet fetch with full-page parsing, PDFs, OCR, scheduled jobs.
- Native connectors to advertising instruments. Slack, HubSpot, GitHub, Notion, Linear, Mailchimp, Resend, SendGrid, Stripe, Gong, WordPress, Airtable, Apify, and even Semrush.
- Skilled talent library. The Ahrefs workforce has contributed pre-built advertising expertise and purposes that encode how we truly work.
Andrei’s essential workflow lives within the GTM Generator, a device that takes one product temporary and produces a whole launch package deal in a single sequence.
For each new product and have launch, Andrei can routinely generate a standalone touchdown web page draft, a 90-second video script, a promotional e-mail, and an nearly print-ready flyer.
There’s then a cross-asset consistency stage that critiques all the outputs and flags any areas the place the message drifts or the main points turn into slightly inconsistent. In observe, that means:
- The video script generator stays capped at 1:30.
- The e-mail targets exactly the correct ICP.
- The touchdown web page covers all the important thing particulars (with out hallucinating something).
Consistency actually issues in PM, so the ultimate evaluation stage additionally reads all 5 outputs side-by-side and writes a abstract.md file itemizing each declare, headline phrase, and ICP framing that disagrees throughout property (like a touchdown web page claiming {that a} new replace was “10x quicker”…regardless of that declare being nowhere within the unique temporary).
As soon as Andrei has edited the property, he may even add the edited file again into the GTM Generator for it run a diff verify between its output and the edited model, permitting it to be taught from the modifications he made.
Starter immediate
Construct me a GTM package deal orchestrator. Enter is one freeform product temporary. Cascade by six levels: (1) Transient: convert uncooked textual content right into a structured contract temporary.md with ICP tiers, positioning, proof factors, KPIs; (2) Touchdown web page: generate a standalone product touchdown web page HTML, hard-rule that bans naming sibling merchandise of the identical firm anyplace in physique copy; (3) Video script: 90-second segment-structured script with VOICEOVER, SCREEN, ON-SCREEN-TEXT per beat, onerous cap 1:50; (4) Electronic mail: single launch e-mail plus 2 topic strains plus 2 preview texts, 120-180 phrases; (5) Flyer: 2-page A4 panorama HTML plus PDF; (6) Consistency verify: LLM reads all 5 property and writes abstract.md itemizing each cross-asset contradiction. Every step writes to a runs desk, per-step rerun supported, the temporary is the upstream supply of reality.
The Touchdown Web page Generator is a part of the GTM workflow above, however Andrei runs it standalone on a regular basis when he need to rewrite a product web page or create a brand new use-case web page.
After you paste a quick in to the device, it pulls related key phrase information utilizing the Key phrases Explorer endpoint, chooses a related template, drafts an overview Andrei can edit, earlier than producing the web page, prepared for hand-off to the online workforce to construct.
Andrei creates a lot of touchdown pages catered to completely different persona varieties (businesses, entrepreneurs, ecommerce, SaaS, entrepreneurs, freelancers, enterprises). The touchdown web page generator has pre-made “seed” templates for every persona, making it simple to generate a number of touchdown variations for every new characteristic or launch.
Starter immediate
Construct me a product touchdown web page generator. Hybrid temporary enter: freeform textarea plus editable structured kind (product title, archetype, ICP tiers, proof factors). Pipeline: (1) competitor scan, learn canonical competitor listing from a talent file; (2) Key phrases Explorer matching_terms, archetype seeds, prime 50 by quantity; (3) define stage with editable part listing; (4) web page technology with a tough STANDALONE_RULE immediate clause itemizing each sibling product of mine, plus a post-generation regex scan that flags any sibling-product title in physique copy; (5) search engine optimisation metadata (slug, meta title, meta description). Output: standalone HTML plus spec markdown plus inline preview.
The Paid Adverts Marketing campaign Builder is what Andrei runs when he need to see how a competitor is messaging of their paid campaigns. The device makes use of the Web site Explorer paid_keywords and paid_pages endpoints to search out any area’s paid campaigns:
It then fetches each paid touchdown web page to search out and analyze the advert copy, clusters paid key phrases into ad-group themes, and offers a strategic overview of the corporate’s paid technique:
Then the generator outputs Google Search creatives for Andrei to evaluation, with three headlines and two descriptions per variant, and three variants per advert group—full with headline, CTA, and full copy:
Starter immediate
Construct me a paid advertisements marketing campaign builder off a competitor’s spend. Enter: competitor area. Pipeline: (1) Ahrefs paid_keywords plus paid_pages on their area; (2) web-fetch prime paid touchdown pages so the LLM reads their advert copy; (3) cluster key phrases into advert teams, flag non-branded, rating hole vs my area (the place I don’t at present bid); (4) generate Google Search creatives, 3 headlines plus 2 descriptions per variant, 3 variants per advert group, onerous char caps (headline 30, description 90); (5) generate 1080×1080 social PNGs in 4 background kinds by way of Pillow with my model font, orange accent phrase. Describe reference advert visuals verbally within the system immediate as an alternative of passing binaries. Persist runs in Postgres, ZIP obtain of all PNGs.
The Battlecard Generator takes a competitor URL and produces a 6-box gross sales battlecard (with sections for Overview, Product, Pricing, Strengths, Weaknesses, Easy methods to place Ahrefs, and The place we’d lose).
The device generates each an HTML preview and an actual .pptx export that mirrors Andrei’s current template:
Sidenote.
That is simply an illustrative instance, and never one thing we truly use.
The battlecard is generated primarily based on tons of analysis:
- Evaluation of core web site pages: homepage, pricing pages, options and options pages, plus to fifteen nav hyperlinks found on the homepage.
- Publicly accessible evaluation websites: like G2, TrustRadius, and Capterra.
- Ahrefs information: like DR, natural visitors, ref-domains DR distribution, prime paid key phrases, branded visitors break up.
- LLM evaluation: information synthesis to reply two essential questions: “Easy methods to place Ahrefs”, and “The place we’d lose”.
Starter immediate
Construct me a gross sales battlecard generator. Enter: competitor URL. Output: a 6-box battlecard (Overview, Product, Pricing, Strengths, Weaknesses, How-to-position-us, The place-we-might-lose) as HTML plus an actual .pptx that mirrors my workforce’s current battlecard template. Pipeline: (1) deep scrape competitor (homepage plus /pricing, /options, /options plus as much as 15 inside nav hyperlinks, cap 18 URLs); (2) evaluation fetch straight from G2 detractor URL (filters[nps_score][]=1), TrustRadius, Capterra by slug, in parallel; (3) search engine optimisation alerts on their area (DR, natural visitors, ref-domains DR distribution, paid key phrases, branded visitors break up); (4) LLM synthesises Strengths, Weaknesses, How-to-position-us, The place-we-might-lose utilizing my technique talent as floor reality for our benefits and ICP framing; (5) HTML render to my workforce’s colour palette plus a python-pptx export that mirrors the present battlecard structure.
The LinkedIn Publish Generator is the smallest device on this listing nevertheless it’s one Andrei and Constance use most days. Paste a tough LinkedIn put up thought, and the appliance returns two polished variations again with two completely different angles to select from:
The generator makes use of a set of predefined “angles” to generate posts: narrative, hook-led, teardown, undertaking reveal. The PM workforce picks two, or the app picks for them.
The device refers to Andrei’s and Constance’s private linkedin-post-bank (a curated collection of their best-performing posts), and makes use of a brand-voice talent to maintain each product point out correct.
Sidenote.
Opposite to the screenshot, I’m not Ahrefs’ Product Advertising lead… however I’ll take a second paycheck if one is going.
Starter immediate
Construct me a LinkedIn put up generator. Enter: a tough draft and two angle picks from {narrative, hook-led, teardown, undertaking reveal}. Learn a private “put up financial institution” talent file as voice floor reality on each technology (it carries 15-30 reference posts in my voice). Detect the sample of the draft, then rewrite into two variations, each fold-safe underneath 1300 chars. Add an “Add to financial institution” button that appends accredited outputs again to the talent file. Persist runs in Postgres for historical past. Reuse the identical financial institution talent file from every other agent session that drafts LinkedIn content material, the file is the corpus.
Constance runs plenty of webinars for Ahrefs, so she requested Agent A to construct the Webinar Automation app: she enters a webinar title, date, and roster as soon as, and it creates a full Linear undertaking from scratch:
The device runs by 5 distinct phases:
- Part 1 (Idea) spins up the Linear undertaking within the Product Advertising workforce with milestones back-calculated from the webinar date—Promo Begin, Reside, Recording Prepared, Report.
- Part 2 (Touchdown) creates a Zoom occasion, Admin touchdown web page, updates the /webinars/ web page, and creates a ticket for the design workforce.
- Part 3 (Promotions) generates promotional copy, an Intercom workflow, and creates a ticket for our paid supervisor to begin promotion.
- Part 4 (Content material) generates drafts of Agent A demos, webinar slides, and a script.
- Part 5 (Reporting) creates a efficiency report back to share with the workforce.
Starter immediate
Construct me a webinar automation Console app backed by Linear. Atomic enter: a webinar title, date, and roster of teammates with their Linear consumer IDs. UI walks 5 sequential phases — Idea, Touchdown pages, Promotions, Content material, Reporting — every with its personal settings panel and a Launch button; part N+1 is locked till part N is launched. Part 1 creates a Linear undertaking in my Product Advertising workforce with 4 milestones back-calculated from the webinar date. Phases 2-5 create 14 points + 6 sub-issues from Jinja-templated descriptions with {webinar.title} interpolation; I can override any title or assignee per-webinar earlier than launch. Templates stay in code so that they model with the app. Use the linear connector for tasks/points and direct GraphQL for milestones.
The Calls to Positioning device is for analysis, not deliverable technology, nevertheless it’s extremely helpful. It takes gross sales name snippets, analyses their content material, and creates ideas for a way the PM workforce can replace our product positioning to win extra offers—all grounded in actual buyer quotes.
Sidenote.
It is a fabricated screenshot much like Andrei’s device, so there’s no PII right here 😉
The evaluation is damaged into 4 classes:
- Why we win: successful patterns and promoting factors value incorporating into web site copy.
- Why we lose or stall: a listing of objections, blockers, churn alerts.
- Verbatim language: phrases prospects truly use, ordered by recurrence.
- Proposed positioning shifts: earlier than/after examples with the supporting quotes hooked up.
Starter immediate
Construct me a sales-calls-to-positioning device. Enter: paste name snippets one per blank-line block, non-obligatory speaker labels. Optionally ship a built-in 15-20 snippet pattern corpus that covers wins, losses, churn, and competitor mentions, so the device is usable on first run. Strip any supply metadata (speaker, stage, deal final result) from the immediate: clustering should come from the language itself, not the tag teams. Output 4 buckets: (1) Why we win, successful patterns with verbatim phrases value lifting into copy; (2) Why we lose or stall, objections, blockers, churn alerts; (3) Verbatim language, phrases prospects truly use, ordered by recurrence rely; (4) Proposed positioning shifts, earlier than/after with the supporting verbatim quotes hooked up. Persist runs in Postgres for historical past.
For Constance, slide decks are the worst a part of each product launch. She used Agent A to construct a Textual content to PowerPoint so she by no means wants to begin from a clean slide once more.
Constance writes her notes and concepts in Markdown and pastes into the device. The device splits the content material into sections and slides, whereas a regex rule catches image-cue patterns (“present screenshot”, “insert diagram”, “[image]”, <screenshot>, “add a chart”…) and reserves a labeled dashed-orange placeholder field on the correct of the slide.
The PowerPoint builder has the Ahrefs palette, and Constance can preview slide-by-slide within the browser, edit any slide inline, and export when she’s completed.
Starter immediate
Construct me a markdown-to-pptx generator. Enter: pasted markdown or .md add. Parser splits on #/##/### into title/part/content material slides. Regex catches 8 image-cue patterns (“present screenshot”, “insert diagram”, “[image]”, <screenshot>, “add a chart”…) and reserves a labeled dashed-border placeholder field on the correct of every affected slide. Claude Sonnet 4.6 rewrites every chunk into 3-5 concise bullets; unique prose is preserved verbatim within the speaker notes (not the slide physique). Use python-pptx with my model palette baked in: navy header bar, orange accent stripe, 16:9, 26pt headings, 20pt bullets. Title and part slides are full-navy with a centered orange accent line. UI reveals a slide-by-slide preview with inline edit + delete per slide. Export as .pptx.
When you’re an Ahrefs buyer, you may attempt Agent A at no cost for one month. Copy any of those prompts right into a contemporary workspace and your personal Agent A will begin constructing the device—or try the appliance library so as to add a few of these instruments on to your personal workspace.
