Worldwide advertising is a job that will get exponentially difficult with every further area and language you add.
For instance, we publish the Ahrefs weblog in eight languages, which implies roughly each significant process—refreshing an previous article, checking hreflang tags, swapping inner hyperlinks, monitoring what a competitor simply shipped—must be achieved eight instances, or eight instances a month, or eight instances per product launch. Fortunately, we now have some unimaginable entrepreneurs on the case.
Erik Sarissky is Head of Worldwide Advertising and marketing & Product Localization at Ahrefs. He runs the worldwide advertising crew at Ahrefs, main a small-but-mighty cadre of worldwide entrepreneurs liable for scaling development in areas like Spain, France, Germany, and Japan.
Takanori Kawaharada is Ahrefs’ Regional Head of Advertising and marketing, Japan. Taka is liable for all of Ahrefs’ advertising channels in Japan: occasions and webinars, content material creation, social media, product advertising, and rather more moreover.
These are a number of the AI instruments Erik and Taka (and the remainder of the worldwide advertising crew) have constructed with our advertising agent, Agent A. Every rationalization features a starter immediate you possibly can paste right into a contemporary workspace to construct your individual model.
What’s Agent A?
Agent A is a advertising agent from Ahrefs—an AI assistant with direct entry to the complete Ahrefs dataset that may perform advertising duties autonomously, somewhat than simply reply questions.
Agent A consists of:
- Unrestricted entry to Ahrefs endpoints. Each endpoint we use to construct Ahrefs is on the market, together with many you can not attain through API or MCP.
- Severe tech stack beneath. Postgres for state, Flask for UIs, an OpenRouter proxy with 300+ fashions, net 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.
- Knowledgeable talent library. The Ahrefs crew has contributed pre-built advertising abilities and functions that encode how we truly work.
The massive one. This device takes one English-language article URL and creates as much as seven publish-ready localized articles, with WordPress shortcodes embedded, inner hyperlinks localized, pictures translated, and a one-click “publish to WordPress” button on the finish.
Earlier than translation, the device pulls keywords_explorer_matching_terms for the article’s major key phrase within the goal nation, then creates a shortlist of localized key phrases to focus on:
Seven languages are presently wired in—FR, ES, JA, DE, IT, KO, ZH-TW—every with their very own tone guides, glossaries, and translation tips.
Each ahrefs.com/… hyperlink within the translated physique will get swapped to the goal language’s equal if a localized model exists, or skipped with a purpose if not. Subdomains (app., assist.) are left alone. The hyperlink map is rediscovered per locale and cached.
Starter immediate
Construct me a multilingual translation pipeline for weblog articles. Enter: an EN article URL (or pasted markdown) + goal language. Supply fetch tries Jina Reader, Ahrefs snapshot, then direct HTTP. search engine marketing step: keywords_explorer_matching_terms for the article’s major key phrase in goal nation, two-LLM curation (one quick mannequin filters, one stronger mannequin ranks), shortlist handed into the interpretation immediate. Translation enforces a per-language readiness gate — refuse to start out if the type information or WordPress shortcode footer for that language isn’t authored but. Rely supply pictures within the immediate and require the identical rely in output with URLs byte-identical, alt textual content translated. After translation, run an internal-link adapter that rewrites each ahrefs.com/… hyperlink to its localized equal (utilizing a per-locale web page map) or skips with purpose. Stream the interpretation reside to the UI. Ultimate stage: one-click publish to WordPress as a draft, plus DOCX export.
With every Ahref weblog publish containing a dozen graphs and movement diagrams, it’s straightforward to see why picture translation turned an actual bottleneck. So Erik and crew constructed a devoted translation device simply for visuals—diagrams, screenshots with annotations, advertising banners.
Add a PNG, JPG, or PDF , decide a goal language and area, optionally write a quick, get again a localized picture:
The 1st step makes use of Gemini to investigate the picture and write detailed localization directions in plain textual content. The device is aware of that Spanish for Spain is totally different from Spanish for Mexico, and Portuguese for Brazil is totally different from Portuguese for Portugal; the mannequin is advised to “adapt cultural parts as wanted for {area}”—forex symbols, instance domains, title conventions all shift accordingly.
Step two passes these directions plus the unique picture into Gemini 3 Professional Picture Preview’s native image-output endpoint, which regenerates the picture with the localization utilized:
Starter immediate
Construct me a picture localizer for advertising visuals (diagrams, annotated screenshots, banners). Enter: a PNG/JPG or PDF (rasterize web page 1), goal language, goal area (e.g. pt → Brazil vs Portugal, zh → China vs Taiwan), optionally available temporary. Two-pass pipeline: (1) a imaginative and prescient mannequin analyzes the picture and writes plain-language localization directions, (2) a native-image-output mannequin regenerates the picture with these directions baked into the immediate, preserving format, model colours, fonts. After the primary render, preserve a chat field open for incremental edits — feed the present rendered PNG (not the unique) plus the correction again into the identical mannequin so format doesn’t drift throughout iterations. Present the evaluation textual content alongside the output so I can see what the mannequin determined to alter earlier than it modified it.
Case research are the highest-effort content material Taka produces: a 60-minute interview, half a day of transcription, a day of writing, two days of revision. So he requested Agent A to compress that to a day.
The Case Examine Generator takes an audio file (mp3 / m4a / wav / mp4, something ffmpeg can learn), reference URLs, photographed handwritten notes, and a basic-info type, and turns it into a cultured case research draft:
The audio is transcribed and speaker labels are assigned to each sentence primarily based on the interviewee type information. OCR is used to extract textual content from photographed pages of handwritten interview notes, reference articles (like different case research) are pulled into the pipeline, and Opus 4.6 creates a case research draft for evaluation and publishing (like this particular case research, revealed on the Japanese weblog):
Starter immediate
Construct me a long-form interview-to-article pipeline. Enter: audio file (mp3/m4a/wav/mp4, any size as much as 60 min), reference URLs, photographed notes (picture/PDF), basic-info type (firm/business/interviewees/theme/goal chars, default 7000). Pipeline: (1) chunk audio into 10-min segments with ffmpeg, transcribe every through openai/gpt-4o-audio-preview, with a tough guard that flips draft to “error” state if each chunk fails; (2) Sonnet 4.6 assigns speaker A/B/C labels utilizing the interviewee type, with the interviewer title auto-filled throughout host turns; (3) Opus 4.6 imaginative and prescient OCRs photographed notes; web-fetch pulls reference URLs; (4) Opus 4.6 drafts at max_tokens=16000 — if output is underneath 85% of goal, run a second “lengthen” cross that provides depth from unused transcript content material however retains the identical title and construction, by no means a part-2 cut up; (5) DOCX + WordPress HTML export, plus chat-refine editor with DB-backed undo/redo. Ban placeholder strings (“[needs answer]”, “[draft]”, any disclaimer textual content) within the system immediate.
Taka ships 4 sorts of movies: month-to-month product roundups, deep-dive demos, step-by-step tutorials, and podcast interviews. Every one has a totally totally different construction, so the YouTube Script Generator has 4 templates to decide on between:
The device additionally features a hook sample picker with six confirmed “hook” codecs to pique the viewers’ curiosity (Unfavorable End result / Social Proof / Sample Interrupt / Excessive Stakes / Pace Run / FOMO Hole). Taka’s selection shapes the primary 30 seconds, whereas the remainder of the script is generated by Opus 4.7 at ~300–350 Japanese characters per minute:
For thumbnails, GPT-5.5 proposes three textual content ideas (an information angle, a query angle, a narrative angle), impressed by a research Taka performed taking a look at thumbnail developments in Japanese business-education channels; gpt-5.4-image-2 renders every at 1280×720 utilizing brand-color palettes.
Starter immediate
Construct me a YouTube script + thumbnail generator with 4 script-type templates: demo (7/12/50/15/8% sections with screen-direction cues each 2-3 strains), month-to-month product roundup (5/80/8% sections, banned hype phrases, exclamation marks solely in outro, mounted opening/transition templates per place), step-by-step tutorial (8/7/65/10/8% sections with specific step numbering), and podcast interview-question script (Host intro + numbered questions, every query = 2-3 sentence premise with particular stats/names + one open query, no visitor solutions, no dialogue). Elective hook-pattern picker for the primary 30s (Unfavorable End result / Social Proof / Sample Interrupt / Excessive Stakes / Pace Run / FOMO Hole). Tempo at ~300-350 [native-language] chars/minute. Fetch reference URLs through web-fetch and inject content material. After script is completed, add a thumbnail tab: GPT-5.5 proposes 3 textual content ideas (information/query/story), picture mannequin generates 1280×720 thumbnails with brand-color palettes pulled from a saved native trend-study file.
For any canonical URL, the tool returns the complete set of hreflang tags (<link rel="alternate" hreflang="https://ahrefs.com/blog/agent-a-for-international-marketing/..." href="https://ahrefs.com/blog/agent-a-for-international-marketing/...">), tailored to the international markets you care about, and ready to paste into a CMS field or a sitemap generator.
Audit mode also allows you to troubleshoot common hreflang issues. Point it at a page that already has hreflang tags and it tells you what’s wrong—missing reciprocal tags, wrong language codes, broken alternates, missing x-default. This catches the half-implemented-by-the-previous-agency state most international sites are in.
Starter prompt
Build me a hreflang generator + auditor for international sites. Support both URL patterns: path-based (ahrefs.com/blog/de/slug) and subdomain-based (de.example.com/blog/slug) via a single slug-prefix parser. Slug matching in order: (1) exact same-slug across languages, (2) Ahrefs Site Audit hreflang_audit endpoint when available (cache to disk), (3) title-similarity fallback. For a canonical URL, output the full <link rel=“alternate” hreflang> tag block including x-default. Audit mode: point at a page that already has hreflang and report missing reciprocals, wrong language codes, broken alternates. Auto-import brand configurations from my Intl Blog Monitor so I don’t reconfigure the same site twice.
Twice a yr Taka hosts or sponsors an occasion the place the speaker is in a single language and half the viewers is within the different. Skilled simultaneous interpreters value $2,000+/day and don’t know our product. The Stay Interpreter is what Taka opens on a laptop computer subsequent to the stage:
This translation device data reside audio and generates subtitles on-the-fly utilizing Gemini 3 Flash (with a selection of different fashions for various conditions: Claude Haiku is extra correct however too gradual for reside occasions).
At Taka’s specification, the subtitling device forces a grammatically full output sentence even when the enter is a mid-phrase chunk of reside speech: no noun-ending sentences, no trailing particles, and should shut with a correct Japanese predicate (the polite-form copula or a verb conjugation).
And crucially from a product perspective: a devoted search engine marketing/Ahrefs glossary is injected on the prime of the system immediate with the literal directive “ABSOLUTE HIGHEST PRIORITY — overrides all the pieces else”. So “AI Overviews” at all times renders because the one accepted Japanese time period we use internally, by no means a near-synonym. “Backlink”, “Area Ranking”, and each different brand-controlled time period map to precisely one Japanese rendering.
Starter immediate
Construct me a real-time convention interpreter for [my two languages]. Browser captures audio chunks, POSTs every transcribed chunk to /translate. Server detects path by character-set / language ID. System immediate enforces two non-negotiables: (1) each output sentence have to be grammatically full with a correct predicate even when the enter is a mid-phrase chunk — if enter cuts off, mannequin provides a closing verb / clause to complete the sentence so subtitles by no means dangle; (2) inject my area glossary (each instructions) on the prime of the system immediate with the literal directive “ABSOLUTE HIGHEST PRIORITY — overrides all the pieces else” so model phrases by no means get substituted with synonyms. Pre-wire 4 mannequin choices ordered by latency, not accuracy: Gemini Flash Lite, Gemini 3 Flash (default), Claude Haiku 4.5, GPT-4.1 Nano. max_tokens=400, temperature=0.2. Present path + mannequin within the UI; swap mannequin on the fly with out reloading.
Ultimate ideas
Ahrefs creates a ton of selling collateral every month, and the worldwide advertising crew has a mountain to climb to localize all of our weblog posts, emails, product movies, YouTube tutorials, and run reside occasions.
Erik and Taka use Agent A to automate large chunks of the less-skilled, extra repetitive work every month. All of those instruments began life as handbook, time-consuming workflows and was an app over a chat session or two. The full work was in all probability a standard week’s price of an engineer’s time, besides that neither Erik nor Taka is an engineer.
When you’re an Ahrefs buyer, Agent A is free to strive for one month. Paste any of the starter prompts above right into a contemporary workspace and your Agent A will construct the device—together with your information, your locales, and your priorities.
