SEO professionals waste hours on repetitive tasks that AI can complete in minutes. We asked 25 practitioners — agency founders, consultants, and in-house strategists — for the exact prompts and workflows they actually use to cut turnaround time. Their answers cover the whole pipeline: research and intent mapping, content briefs, technical SEO automation, winning citations in AI search, and link building.
A pattern runs through almost every one of them: the highest-leverage AI work in SEO isn’t generation, it’s structure — feeding the model real project data, forcing a fixed output shape, and keeping human judgment on the final call. Below is the full roundup, grouped into five themes. Each theme also has a dedicated deep-dive guide linked at the top of its section.
1. AI for SEO research & planning
Deep dive: How to use AI for keyword research, search intent & content planning »
Map search intent before any drafts
The workflow that has saved me the most time is an AI-assisted search-intent map before any page gets written or revised. I use Codex/ChatGPT plus search data, Search Console exports and the live SERP. The prompt structure is: “Act as an SEO analyst. Cluster these queries by intent, identify the page type Google is rewarding, note which questions AI Overviews or answer engines are likely to summarize, then give me the content gaps, schema opportunities and internal-link targets.”
The key is that I do not ask it to write the page first. I ask it to diagnose the search pattern. That turns a messy keyword list into a working brief much faster: primary intent, secondary questions, comparison angles, objections, entity terms, and what needs proof. Then a human decides the claim, examples and offer. The blank-page phase disappears.
Build a master entity table first
The process that helped me turn around faster is building a single entity table before writing a single page. I feed the business name, list of services and address into the system and prompt it in a specific way — asking it to include specific entities, services and classes of related topics in a table, with columns showing the entity, its importance and page location. I get one table that forms the template of the entire website so nothing important is missed.
The prompt structure is the one thing that works, no matter which system I use: I provide three elements in a specific order — the role, the business details, and the required output format. Vague prompts yield vague responses. The unique step is bulleting that table into questions for each page; because it’s generated from a structure I already approved, the result stays consistent across 30–60 pages instead of going off in different directions.
Convert strategy into a content taxonomy
One workflow that reduced my turnaround is using a custom GPT to convert a marketing strategy into a full content taxonomy. It doesn’t create articles — it organises topics into content pillars, topic clusters, and articles. First I give the business objectives, target audiences, products/services, and priority keywords. Then I tell the model to build a taxonomy following the hierarchy without duplicating topics: Strategy > Content Pillars > Topic Clusters > Articles.
This used to take several hours on spreadsheets. Now it takes 20–30 minutes, and it lets my team map keywords and internal links while writing content with a proper structure.
Reverse-engineer Perplexity’s source preferences
My favorite hack is using Perplexity to research its own prompts. Choose a buying question and put it in 8–10 variations. Paste all the sources it finds into a spreadsheet, and after ~15 minutes you’ll see what pages it consistently pulls for each version of your query. Those are your templates. Study their formats — lists? numerical bullet points? — that’s what the AI is prioritizing.
Make sure to incentivize the model to disclose this too. I always add “list your sources with URL” and “justify why you picked each source” to my prompts. The latter tricks the bot into explaining its process (“it had an excellent comparison chart” or “it was the most recent article”). Hours of manual SERP research compress into ~40 minutes. Sources shift weekly, so I re-run analyses monthly.
Assign a model the job of SEO analyst
Early AI experiments gave sophisticated answers that were essentially useless — it wasn’t the technology, it was the data going in. What made the difference is having ChatGPT or Claude perform the job of an SEO analyst before writing anything. We give it real project data: exports from pages, keywords, search intent, titles, headers, links, and conversion targets.
The instructions are simple: state the business goal, provide the page data, explain the target audience, then ask the AI to identify topics to cover, page duplication, competing pages, and whether a merge or rewrite is needed. Audits that took a few hours now reach first-draft stage in half the time. The moral: don’t ask AI for SEO recommendations — tell it what the project is about and give it everything you have. Treat it like a junior analyst who knows nothing about the client.
2. Client context & content briefs
Deep dive: How to brief AI for SEO content (client context that kills hallucinated links) »
Centralize client context for better briefs
The biggest shift wasn’t a tool — it was deciding to stop starting from scratch every time. Instead of re-explaining the whole client in every new chat, we build a dedicated Claude Project per client and load it once with everything the model would otherwise guess at: brand voice guide, ICP and sales-call objections, a full URL inventory (every page with target keyword and current position), the internal link map, SERP notes, and CMS constraints.
The prompt itself is four parts and the order matters: (1) make it recall the constraints first — “before you answer, pull the three pages from the URL inventory closest in intent to [keyword] and tell me whether to build new or consolidate”; (2) the task; (3) the output format, spelled out; (4) make it check its own work against the inventory and flag any anchor pointing at a URL that doesn’t exist. Steps 1 and 4 do most of the work — the last one catches hallucinated internal links, the single biggest reason teams try AI for on-page work, get burned once, and never go back. Content briefs went from ~3 hours to ~25 minutes.
Configure client-specific rules before outlines
We create dedicated chats for each client so the thread stays focused on one product or service. We set up the initial prompt by giving it the background of an SEO professional with 20 years’ experience in our country, then add rules: “don’t use em dashes,” “write in UK English,” “write professionally,” “use relevant industry jargon.”
Next we paste a list of URLs from the client’s site for internal linking (minimum 5 internal links, at least 1 to another related blog article), require at least 2 outbound links to supporting industry or government sites, give it the primary keyword and ask for at least 4 keyword variants at 1–2% density, have it create the schema code, and generate a 60-character SEO title and 160-character meta description. Then we review and manually edit every article.
Combine SEMrush and guidelines for content plans
One of our most effective workflows matches the best of AI with the best of humans. A specific prompt — combined with Claude connected to SEMrush and equipped with the client’s content and tone-of-voice guidelines — produces four blog plans that target actually-searched keywords from H1 down to H3, plus trending topics.
The prompt: “Please create four blog plans for [client] ([website]). Each should include a suggested H1 and at least 5 H2 subheadings, all in one document. The goal is helpful advice targeting a common question or topic (use SEMrush keyword research). Make the H1 target the main commonly-asked question and the H2s target related questions under it. Include, for each: keywords targeted and their search volumes, suggested H1, suggested H2s, and bullet points under each subheading. Target market: [demographic and key problem].”
Mine sales calls for real objections
The workflow that moved our turnaround has nothing clever in the prompt. We feed the model twenty to thirty of the client’s own sales-call transcripts before we ask it anything about content. Then the prompt runs one direction only: list every question a prospect asked in these calls, group the ones asking the same thing in different words, and rank the groups by how often the objection appears. We never ask it for keywords.
The output is a list of real buyer language nobody else in the SERP has, and we map that against search volume afterward rather than starting there. Brief production dropped from roughly two days to under two hours per page. The generic version fails because the model only knows what’s already published — so it hands you back the same page everyone else is writing.
Standardize pre-sales audits with a reusable skill
The workflow with the biggest impact for me isn’t a prompt, it’s a reusable Claude Skill that takes a prospect’s URL and produces a branded search-visibility audit in minutes. The same audit used to take 2–3 hours manually; I’ve now used it for around 30 prospect audits and it’s become the core of my lead generation.
The only variable is the website URL. Everything else stays fixed: a predefined scoring rubric based on my three-layer Search Visibility Framework, consistent section order, UK English, fixed brand voice, formatting rules, and strict instructions on what not to do (invent metrics, make unsupported claims, use generic advice). The workflow separates retrieval from reasoning — live information is gathered first, then the model evaluates that evidence against my framework rather than generating an opinion from scratch. I still review every audit manually; anyone sending AI-generated audits straight to clients without that review is taking a risk.
3. Winning citations in AI search (AEO/GEO)
Deep dive: Answer Engine Optimization: how to get cited by ChatGPT, Perplexity & AI Overviews »
Lead with answers to earn citations
One workflow that genuinely cut turnaround: AEO restructuring. Take a client page that’s ranking but not getting AI citations, figure out what question it actually answers, and rewrite the intro to lead with that answer in sentence one. Used to take 45+ minutes per page.
The prompt: “List the 5 questions a user might have that this page answers. For each, write a 2-sentence opening that answers it directly and could stand alone as a cited excerpt.” Pick the best fit, drop it in — under 10 minutes. The old versions were SEO-optimized but buried the answer in paragraph 3 after the obligatory intro. Restructuring to lead with the answer is what gets you cited. Citation is the new click. The bottleneck shifted from writing to judgment — deciding which question each page should own.
Score buyer questions across response engines
The workflow that changed my turnaround is ten fixed buyer questions per company, run across ChatGPT, Claude, Perplexity and Google AI Overviews, with every answer scored twice — once for whether the company is named, once for which domains got cited as evidence.
The prompt structure is the part that matters. I never ask an engine to rate anyone; I ask the question a buyer would actually type, then score whatever comes back. Asking a model to evaluate a vendor produces flattery. Asking it to answer a buyer’s question produces data. That turned a day of reading answers by hand into a scored dataset I can compare month over month.
Filter listicles to target eligible pages
Listicle prospecting works because it never asks the model to be creative — it asks it to filter. What AI assistants quote back is overwhelmingly “best X” list pages, so getting onto those pages is a distribution problem worth automating.
Stage one is pure search: dozens of Google SERP queries built from the category and its modifiers (“best X”, “X alternatives”, “top X tools”), returning ~1,264 results per run. Stage two is where the model earns its place: every result goes in with just the title, URL and meta description, and one instruction — decide whether this is a list a new tool could realistically be added to. Answer in one word plus a reason under ten words. Around 33 survive. Stage three pulls the stated ranking criteria from each of those so the pitch references what the writer cared about. The whole trick is one decision per call, a forced output shape, and a hard cap on reasoning length.
Reveal prompt gaps and brand mentions
GroundScore.ai is a tool I built for my own SEO/AEO agency to spot gaps and auto-generate content to fill them. It shows what people ask in AI prompts about your business, where you’re mentioned in AI answers, generates content around the prompts where you’re not cited, and shows where competitors appear in areas you don’t.
4. Technical SEO automation
Deep dive: Automating technical SEO with AI — schema, cannibalization & prompt pipelines »
Detect and fix keyword cannibalization
Keyword cannibalization happens when two or more pages compete for the same term, pulling each other down. Normally, finding these takes hours of cross-referencing GSC data in Excel — this workflow does it in under 10 minutes.
Export the last 3–6 months of query and page data from GSC as a CSV, upload it to ChatGPT (with Advanced Data Analysis) or Claude, and run: “Act as a Lead Technical SEO Strategist. Analyse the attached GSC CSV for keyword cannibalisation. 1. Find queries where 2+ different URLs get clicks or impressions. 2. For each, identify the Primary Page (most clicks/impressions) and the Competing Page. 3. Recommend a specific fix — 301 Redirect (if the competing page is outdated), Canonical Tag (if both stay live), or Internal Link Fix (give the exact anchor text). Format the output as a clear table.” No manual VLOOKUPs, and it tells you exactly what to fix.
Automate a consolidated JSON-LD graph
The workflow that moved the needle for me is automating deeply nested JSON-LD schema. Generic CMS plugins don’t cut it — to feed clean entity data to AI search engines the schema needs to be deeply connected. Instead of asking AI for one schema type at a time, I feed it the scraped text of a landing page and use a “Master Graph Prompt” so it links the page’s entities back to the main canonical URL using stable @id fragments.
The prompt: “Generate a unified page-level JSON-LD script. Create a single JSON-LD block with a @graph array (not multiple script tags). Generate a WebPage node with @id ‘[Canonical URL]#webpage’, a Service or LocalBusiness node with @id ‘[Canonical URL]#service’, and an FAQPage node if there are Q&As in the text. Link all page-level entities back to the site-wide Organization @id. Strict JSON — no trailing commas, no smart quotes, no markdown outside the code block.”
Run a prompted pipeline in code
The workflow that actually changed my turnaround is running my whole content process inside Claude Code, with every prompt saved as a file in the project instead of typed into a chat. I have a research command that pulls live keyword volumes and SERP data through the DataForSEO and Ahrefs APIs, a write command that drafts against that brief, and a publish command that commits the finished article straight into my static site’s git repo.
Each command file spells out the role, which data sources to call, the exact step order, and the output format. Because it lives in git, a bad draft means I fix the prompt file once and that failure mode is gone for good. Nothing ships on vibes — every draft has to pass deterministic scripts first (a scrub for invisible Unicode watermarks and an AI-detection scan). A research-backed article used to take about two working days; now it’s closer to two hours.
Connect MCP servers for integrated research
Connecting MCP servers to Claude or ChatGPT has been one of the biggest time-savers in my SEO workflow. I’ve connected the official Ahrefs, GA4, and Google Ads MCP servers so I can ask questions or do research straight from the chat, with more insight than I’d get doing it all manually. It also helps analyze competitors and build proposals for prospective clients. It can’t be fully automated yet, but it improves the end result and saves a lot of time.
5. Content QA, optimization & link building
Catch self-contradictions before publication
The one that changed our turnaround isn’t a content prompt — it’s a contradiction check against our own archive, run before anything gets written. Our published archive and internal decision notes sit in a vector index. Before drafting, we query it semantically with the thesis of the new piece, and the prompt is deliberately adversarial:
“Here is the claim I intend to make: [claim]. Below are excerpts from our own published work. Identify anywhere we have stated something that contradicts this claim, anywhere we made a weaker or stronger version of it, and any commitment this would violate. Quote the exact sentence and give the source. If nothing contradicts it, say so plainly rather than finding something.”
That last instruction matters most — without it you get manufactured conflicts. Research went from half a day to under an hour, and our rate of publicly contradicting our own earlier positions went to near zero. Index your decision notes, not just published pages, and keep the query semantic.
Prioritize rewrites from GSC data
I use Claude to turn raw Google Search Console export data into a prioritized action list — not to write content at all. Every week I export queries and pages (impressions, clicks, position), filter to anything ranking between position 4 and 20, and drop the CSV into a fixed prompt: (1) group rows by page URL and sum impressions; (2) flag pages above an impression threshold with average position 4–20; (3) for each, pull the actual query text driving impressions and check whether the current title and H1 target that query directly or only tangentially; (4) output one specific rewrite suggestion per flagged page.
What took half a day of scanning spreadsheets now takes twenty minutes. Forcing the model to compare query text against the actual existing title tag is what made it specific instead of vague.
Humanize press releases with deterministic QA
I built a four-stage press-release humanization workflow in n8n that cut turnaround from three days to six hours while meeting the zero-AI-detection benchmarks clients demand. Four prompts run in sequence through the Claude API: stage one scores the draft against 18 AI trigger patterns (sentence-rhythm uniformity, opening-phrase predictability, hedge-word density, three-item parallel construction, passive-voice ratio); stage two rewrites only the flagged sections with constraints enforced in the prompt; stage three rescores with the same rubric and loops back if it drops below threshold; stage four runs the final text through detection APIs in parallel.
The biggest shift wasn’t speed, it was predictability — the workflow enforces the same quality bar regardless of who drafts, and that consistency translated directly into higher client retention.
Produce client-ready wireframes from analysis
We take a new client from keyword research to wireframes they can react to in a day instead of weeks. Research stays partly manual: I use Google and Ahrefs to find competitors ranking for our target keywords, with a bias toward low-competition bottom-of-funnel terms, plus any sites and palettes the client likes. That goes into Claude with roughly: “Review these competitor sites and this keyword data. Group everything into page clusters, and for each cluster give me the primary keyword, the search intent, and the sections the page needs. Then produce a wireframe of the new site built on that structure, designed for [brand] with the styling provided.”
Wireframes used to take weeks of back-and-forth. Now I generate several versions in minutes and put them in front of the client — when people see something physical, they figure out what they actually wanted, fast. It’s a big part of why we could cut our base rate on standard builds by almost half.
Transform reports into actionable task lists
The real advantage for me is processing SEO data with AI much faster. My process starts by getting an SEO report from SeoSets, then feeding it to my AI software. I ask it to sort the findings into categories, focus on the problems that need addressing, and create an action list — and to differentiate between findings and recommendations based on the data. The prompt is simple; the power comes from giving the AI real SEO data and asking it to do one specific thing. Then I analyze the results against the web page, search intent, and business goal. The most efficient use of AI for SEO is getting a concise action list out of unstructured data.
Turn Qwoted opportunities into link assets
A Claude Code skill turned Qwoted pitching into a genuine backlink pipeline. Rather than sending a plain opinion pitch, the skill researches the topic and builds a fully sourced statistics page — with charts and structured data — as a linkable asset to pitch alongside the quote. That page becomes the thing journalists actually want to cite, for months afterward.
For a SaaS client, Claude built a statistics page pulling together dozens of sourced data points on pricing trends, published it, then pitched three relevant journalist requests linking to it. Two landed coverage within the week with do-follow links, and the page picked up two further citations organically over the following month. What used to take a full day of searching and writing now happens in under an hour.
Find broken-link prospects for backlinks
A simple AI prompt can put the information you need right at your fingertips. I use it to find wellness blogs with broken links in our area, then tap into their traffic for leads. I’ll type something like “list the top ten health and wellness websites in the USA that have broken links,” check that each site aligns with our products, then contact the owners to offer new content in exchange for a backlink. Since going this route, traffic jumped by more than 1000% (other strategies contributed too).
The common thread
Across all 25, the productivity gain almost never comes from asking AI to write. It comes from three moves repeated in different forms: feed the model real data (transcripts, GSC exports, entity tables, your own archive) instead of asking it to invent; force a fixed output shape (one decision per call, a table, a rubric, a strict JSON block); and keep the final judgment human — deciding which question a page should own, verifying claims, and adding the commercial context only experience provides. Structure is the leverage. Generation is the easy part.
Want help turning any of these into a repeatable system for your site? Salam Experts builds technical SEO, content, and AI-search workflows for clients across surety, transport, local services, and international markets. Get in touch »
