The offsite playbook

Buyers moved faster than most teams

SEO teams have spent decades controlling the one thing they could control: their own website. For traditional search that was the whole game. AI search broke that, and buyers moved faster than most teams did.

These models decide which brands come up when someone asks the questions that actually matter:

They read your website, then they read everything else, and the everything else does most of the work. If the sources they trust don't mention you, the model has nothing confident to say about you.

Two studies make the case

So your own site only takes you so far. The rest of it lives in the sources these models cite, and you need to be in enough of them for it to hold.

Time

Roughly ten hours per mention, averaged across the whole process. Four mentions is the floor, so budget around 40 hours to start. Half or more of that lands before you send a single email.

Setup runs 20 to 25 hours: building the prompt list, running every prompt across every model, expanding through Google, and scoring what comes back. A 25-prompt list surfaces around 1,000 sources and you throw most of them away, but you have to look at them first to know which ones.

Most of those hours go to running prompts by hand, and a visibility tool takes that off you. The ones that stay are spent writing and editing the prompt list, and those decide how good every step after it is.

Money

Three costs.

What success looks like

Those ranges assume replies are coming and publishers are moving. Starting cold, with no relationships and no reputation with editors, puts you at the back of all of them.

At month three, count live mentions. The AI answers move later, and four is the number to count toward. Everything before that is still setup.

Still deciding whether this is worth doing at all? We put the evidence in a separate piece: Is AEO worth investing in.

What this guide includes

You're optimizing for questions now. Whole ones, phrased the way people actually type them into an LLM.

Before you start, know what this list is worth. Roughly 40 citations surface per prompt, so a 25-prompt list surfaces around 1,000 sources. That number is higher than it used to be, because each prompt now gets run repeatedly across ChatGPT, Claude, Grok, DeepSeek, Perplexity, AI Mode and AI Overviews, and each run will likely return a different set. See study from SparkToro on this. The list you build here sets the size and shape of every step that follows.

How to do it

1. 1

   Open a doc or spreadsheet.

2. 2

   Brainstorm the main money questions people ask in your category, for example:

   • –"Best [category] tools for [audience]"
   • –"How do I [do the thing your product solves]?"
   • –"[Competitor] alternatives"
3. 3

   For each question, write 3 to 5 variants. Change the phrasing. Try a short version and a long one, like "best project management tools" against "which project management tool should I use for a remote team."

   When you write those variants, change functionality and persona first. Those two produce the biggest swing in which sources surface. "Help desks for SMBs" pulls a different citation set than "help desk software."

4. 4

   Add columns: Question, Intent, and Priority from 1 to 3, where 1 is critical for revenue or positioning.

Leave branded prompts off the list. They tell you what AI already says about you, which is a different job. The one exception is "[Your Brand] vs [Competitor]," which is high intent and surfaces sources you want to be in.

Use a tool for the first draft

Any of the visibility tools will draft a prompt list for you, and they handle the running in Step 2 as well. Conductor, Otterly, Profound and Scrunch all do this.

Then curate it hard. A list carried over from a tool unedited is a list of prompts nobody chose. Read every one before you commit to it.

A weak list quietly wastes the next four steps, because you'll spend them scoring and chasing sources that were never going to matter. A few things to check as you review your list:

1. 1

   Cut the duplicates. "What is treasury management?" and "How does treasury management work?" lead to the same answer, so keep the clearer version and drop the rest.

2. 2

   Drop anything off-target. If a prompt doesn't reflect what you actually sell, it points you at the wrong sources.

3. 3

   Prioritize by intent. AI prompts don't come with search-volume data, so the useful question is whether the answer puts you in front of the right buyer.

If you want to go deeper on building your prompt list, we wrote it up separately. And for the exact prompt you can plug into an LLM to start building a thorough list, copy the template here.

Two passes over the same list. First you capture what the models already cite. Then you widen it with what Google ranks, because the models pull from the same ecosystem.

Pass one: what the models cite

1. 1

   For each question from Step 1, run it in ChatGPT, Perplexity, Claude, Gemini, AI Mode and AI Overviews. Cover as many as you can stand. Each one surfaces citations differently.

2. 2

   Run each question more than once. Answers personalize per user, so a single run is a sample of one. Use different sessions or accounts where you can, and treat what comes back as a distribution. Every buyer gets a different AI answer

3. 3

   For each response, open the Citations tab and fill one row per cited URL: Target Question, Model, Brand Mentioned?, Competitors Mentioned, Cited URL, Citation Type, and Date Run. Set Discovery Source to LLM.

4. 4

   Collect now, de-duplicate later.

Pass two: what Google ranks

The models don't only draw on the exact pages they cite. They draw on the same ecosystem Google shows humans, so the pages ranking for your question are often the next ones to get cited.

1. 5

   Take one priority question and Google it in an incognito window.

2. 6

   Work through the top 10 to 20 organic results, applying the same filter you'd apply to anything from pass one.

3. 7

   Add each keeper to the same tab, with Discovery Source set to Google SERP and the SERP Position filled in.

Repeat for every priority question.

What to keep, in both passes

Keep guides and explainers only when they're already close to making a recommendation. They're what most brands chase first and they yield the least.

Ignore homepages, pure product landing pages, and obvious junk or AI spam. This applies to pass one too. A model citing someone's homepage is still a citation, and there's no version of this where you get a mention added to it.

Check the date before you add anything.

Between two similar targets, take the one updated most recently. A four-year-old listicle nobody has touched is a weaker target than a thinner page refreshed last quarter.

If you're running a visibility tool

It exports pass one for you, which is why this is where the manual method breaks down first. Pass two you'll still do by hand. If you're still shopping, we wrote up the main options here and the underrated ones here.

You now have a Frankenstein sheet of URLs. Time to work out which ones are worth chasing.

How to do it

1. 1

   De-duplicate by URL using your spreadsheet's Remove Duplicates on the URL column.

2. 2

   For every URL left, open it and skim. Is it actually about the question? Does it mention you? Does it mention competitors?

3. 3

   Fill in the Scoring tab: Already Cited by LLM?, Prompt Count, Top SERP?, Page Type, Brand Mentioned? and Competitor Mentions. Use the same Page Type labels as Step 2: comparison, listicle, review, guide, other.

4. 4

   Assign a priority score:

   3cited by an LLM and ranking on page 1 or 2

   2one or the other

   1relevant but neither

   Filter to 2 and above. That's your hunting ground.

5. 5

   Sort what's left by Prompt Count, highest first. A page cited across eight of your prompts is worth several cited once.

Now classify what survived

Only the pages that cleared the filter are worth classifying, which saves you the work on everything you just discarded.

A roundup site that sells affiliate placements and a newsroom with an editorial firewall look identical in a list of URLs. They are nothing alike as a pitch, and sending the wrong one to a newsroom costs you that relationship permanently. This part isn't hard. It's just easy to skip.

Open the publisher's homepage and ask three questions in order. Stop at the first yes.

Mark these Skip

No pitch works on these and some of them will react badly:

Retailer buying guides are different, and worth treating as commercial publishers. The product listing pages themselves aren't worth pitching.

Why prompt count beats domain rating

How often a source gets cited for the questions you care about is the best predictor you have that a mention there will move anything. This is the step where SEO instinct works against you, so it's worth being blunt about the numbers.

In other words, the brands AI names are the ones other sites talk about, more than the ones with the strongest link profiles.

A Semrush study of 1,000 domains found nofollow and dofollow links correlate with AI mentions almost identically, 0.509 against 0.504. The link attribute link builders negotiate hardest over showed no difference in this data. That points to the mention itself doing the work, more than the link attached to it.

So the rule that follows is our judgement based on what the research suggests:

The first is already being cited for your questions. Use domain rating to break ties between otherwise equal targets.

You know which pages matter and what kind of publisher is behind each one. Now you need a name.

How to do it

For each URL scoring 2 or above that you didn't mark Skip:

1. 1

   Open the article and look for a byline, an author bio, and a contact or about page.

2. 2

   If you find an author, search their name plus the site name on LinkedIn, or use Hunter, Apollo or RocketReach to find an email.

3. 3

   If there's no author, use the titles below to find the right role at that company on LinkedIn.

4. 4

   Fill the Contacts tab: Author Name, Author Byline URL, Contact Role, Contact Email, LinkedIn URL, Company and Confidence.

Who to ask for, by publisher type

Commercial publisher

Head of Commerce, Director of Affiliate, Commerce Editor, Head of Partnerships, Business Development. At larger publishers the commerce team owns the roundups as a revenue line, and they're the ones who can say yes.

Independent site

Founder, Owner, Publisher, Head of Content. Usually the same person.

Brand blog

Head of Content, Content Marketing Manager, Head of SEO, Organic Growth.

Newsroom

Standards Editor, Corrections Editor, Managing Editor, or the section editor whose beat matches the article. Many newsrooms publish a corrections address directly, so check the masthead before you look anywhere else.

Most of them won't have a findable email

That's the normal shape of this work, so budget for it. Contact forms convert worse and they still convert, and on a list of forty domains the difference between reaching thirty and reaching eight is the whole project.

How big this list needs to be

Four live mentions is the floor, and you get there with fewer targets worked harder.

Two things cap your list. That one-in-five email rate, and the fact that the pitch which works is a drafted insertion written for one specific article, which costs real time per target. So the working list runs to dozens.

This is also where the ten hours per mention goes.

Our own reply rates run 1 to 2%, with infrastructure built for it. From a cold domain, expect less.

Whether your category has enough

You already have this answer. Count the rows scoring 2 or above from Step 3. That number is your inventory, and most people don't ask the question until they're three months in.

If it's thin, the fix is usually upstream. Go back to Step 1, broaden the prompt list, and re-run Step 2. If broadening doesn't move it much, that tells you something real about your category. Some have a deep bench of comparison and review content. Some don't, and knowing that in week two is worth more than finding out in month four.

What a good mention looks like

It all comes down to these four things. Every ask below is a page failing one of those.

Start with the errors

Read the pages that already say something about you first. You marked them in Step 3 under Brand Mentioned, either Yes or Weakly. That's your shortlist, already built.

You're looking for anything wrong. A stale price. A feature you shipped two years ago described as still missing. A description of a product you stopped being. A category label that was right in 2024.

The publisher has an editorial reason to say yes. An addition asks them for a favor. A correction hands them a more accurate page, which is the thing they already care about.

Most people skip this and go straight to asking for additions. That's why corrections are where the easy wins sit.

Then the gaps

Pages that answer a question you belong in and don't mention you at all. This is the ask most people start with, and it's the one that takes the most work to land, because you're asking for something the publisher hasn't already committed to.

Then the weak placements

You're on the page, accurately, but buried at ten of twelve, or sitting beside companies you don't belong next to, or described in words that work against your positioning. All three are upgrade asks.

Now write the copy

This is the single biggest lever on whether anyone says yes.

Don't send an email asking to be included and leave the publisher to work out how. Send the drafted insertion. Write the sentence or paragraph exactly as you want it to appear, in their voice, matched to the structure of the piece.

Then give the reason it belongs there. Name the specific thing your product does that the article's framing calls for, or the data point the piece is missing. "We're a great fit" is not a reason.

Two things make this work. It removes the labor, which is the real objection most of the time. And it turns your email from a favor request into an editorial decision someone can make in thirty seconds.

If you have proprietary data, benchmarks or original research that would genuinely improve the article, lead with that. A publisher who gets a better piece out of the exchange has a reason to say yes that has nothing to do with you.

Paste the drafted copy into Drafted Copy in the Contacts tab. That's what Step 6 sends.

Write your guidelines once

You'll do this dozens of times, and starting from a blank page every time is how the attributes drift. We put together a brand guidelines document so you're working from one source instead of rewriting your positioning per publisher. How to write brand guidelines

You have drafted copy sitting against every row. Now it goes out.

Two things to settle first

Most rows won't have an email

Roughly one in five publisher domains gives you an individual email address. The rest need a contact form, a generic inbox, or LinkedIn. Forms convert worse and they still convert, and on a list of forty domains the difference between reaching thirty and reaching eight is the whole project.

Log which one you used under Contact Method.

The templates

Keep them short. The drafted copy from Step 5 is doing the persuading. All of them are here.

Corrections go first, and that template works for every publisher type. Attach no offer to it. The power of a correction is that it's editorially justified, and money muddies that. Then one addition template per bucket: commercial publisher, independent site, brand blog.

Remember that although templates save time, you're still speaking to a human. Taking a moment to read the article and show genuine interest in the work a publisher puts in will go a long way. A spam-cannon approach to brand mentions won't work.

Sending

Export your Contacts tab to CSV and load it into Smartlead, Apollo, Mailshake or similar. Tokens you'll want: first name, site name, article title, and the drafted copy itself.

Cadence: initial email, then one or two follow-ups over 7 to 14 days. Stop after that. Log the dates in the Outreach tab.

The form and LinkedIn rows don't go through a sending tool, and they're four out of five of your list. Work those by hand, a few a day, pasting the same body into the form or the message. It's slow and it's most of your reach.

Once the drafting gets quick, the temptation is to generate forty emails and fire them all at once. Resist it. This works because each pitch is specific to an article someone actually wrote. Volume turns it into the thing publishers already delete from their inbox without a second thought.

Some will ignore you. Some will say no. A few will say "sure, but."

Sort the replies

Hard no

“We don't do this.” Mark it and move on.

Soft no

“Not sure it fits.” Worth one reply if you can answer the objection specifically.

Info request

“Tell me more about your product.” Send the detail, keep it short.

Yes

They'll add it.

Counter

Either a rate, or an offer to add you if you add them.

Decide your boundaries before the replies land

Three questions, much easier to answer now than mid-thread:

1. 1

   What's your budget per placement, if any?

2. 2

   Will you mention publishers on your own site in exchange?

3. 3

   What won't you accept, on brand or compliance grounds?

Three deal types

Log the outcome under Deal Type in the Outreach tab.

Is this link buying?

Someone will ask, and it'll probably be your own SEO lead. Structure it so the question doesn't arise.

Google's link spam policy concerns links that pass ranking signals. What you want here is the brand mention.

So you lose basically nothing by asking for a nofollow or sponsored attribute, or by taking the mention with no link at all.

The defensible position is the one that keeps you clear of link spam anyway: placements on pages that are useful to their readers, with the commercial relationship disclosed wherever it needs to be. Ask for that up front. It removes the objection and costs you nothing that matters.

If you're paying, handle disclosure properly

This section is written to US rules. If you're operating elsewhere, check your own regulator.

Paying a publisher to add a mention is a material connection, and in the US the FTC expects material connections to be disclosed. That obligation sits with the publisher. You shouldn't be the reason they skip it.

Let them apply their own policy. Most publishers who take paid placements already have a disclosure standard. Ask what it is and follow it. If they don't have one, suggest they label it.

Only pay for placements that are true. The payment buys the update. It doesn't buy a recommendation you haven't earned. If your product doesn't belong in a roundup, paying to sit there is a disclosure problem and a bad placement at the same time, because readers and models both notice when a mention doesn't fit the text around it.

The version that works long term is a paid update to an article that's more accurate afterward than it was before. Anything else is buying a liability.

Once they've agreed, send a clean packet

They already have your drafted copy from Step 5. What they need now is everything required to publish without another email:

• The final copy, revised if the conversation changed it
• The URL and the anchor text you want
• Exactly where it fits: "after [Competitor X] in the list," or "as a bullet under [section]"
• A short alternate version, in case they want to adjust

Check the attributes one last time before it goes. This is the last point at which you control the words.

If you agreed to anything that changes your own site, you have to deliver. The publisher is usually waiting on you before they publish theirs.

This matters more than it looks

Reverse mentions are the only offer that works with brand blogs, which is one of the four publisher types you'll be working. A meaningful share of your yeses depend on you holding up your end. Miss them and you lose the placement and the relationship in the same week.

How to do it

1. 1

   Filter the Outreach tab to Deal Type = Reverse Mention. That's your obligation list, already built.

2. 2

   For each one, record where on your site the mention goes, how you agreed to describe them, and any deadline you committed to.

3. 3

   Draft the copy. Same standard as Step 5. It has to be accurate, sit sensibly in the page, and read like it belongs there.

4. 4

   File a ticket with whoever controls your CMS, or ship it yourself if you can.

5. 5

   Check the live page afterward. The mention exists, the link works, the copy survived the edit.

6. 6

   Send a short note confirming you're done. This is often the thing that unblocks their side.

These are standing obligations

You've agreed that a mention will be on that page. If someone restructures it six months from now and the mention disappears, you've quietly broken the deal and you probably won't find out until you need that publisher again.

Note the page and the commitment somewhere your content team will see it before the next redesign.

A mention doesn't count until it's live on the page. A live mention doesn't show up in AI answers straight away.

The timeline, because this is where people quit

1. Month 1Setup and outreachNothing live, nothing to measure.
2. Month 2Offers come backPublishers take 30 to 90 days to publish once they've agreed.
3. Month 3First mentions go liveModels pick up a new mention within about two weeks of publication.Count live mentions
4. Months 4–6Volume builds toward fourA readable change in visibility lands in this window.Visibility reads

That's the most common way this gets abandoned. Those ranges also assume replies are coming. If not, this process becomes much longer.

Verify each mention

Open every one. Editors trim, and an inserted sentence is an easy cut.

1. 1

   Open the Verification tab and work through every agreed placement.

2. 2

   Open each URL. Confirm the copy is there, it reads the way you agreed, and any link points where it should.

3. 3

   Set Placement Status to Live, or Broken if the copy changed in editing.

When to check for visibility change

Monthly, starting month three. Re-run your priority prompts from Step 1 exactly as you ran them in Step 2, and log the date.

Run each prompt more than once, same as before. As we've already discussed, a single run tells you almost nothing.

What to measure

• Use your live mention count to compare against the benchmark. Four is the floor, and according to our research, brands that landed fewer than four rarely held whatever lift they saw.
• Presence in answers for your priority prompts, as a share of runs.
• Citations pointing at pages where you now appear.

What not to expect

Click attribution will understate this badly. A buyer reads an AI answer, absorbs that you exist, and arrives weeks later through a branded search or direct. 68% of US Google searches already end without a click, per SparkToro and Similarweb, and AI answers behave the same way. Measuring sessions from ChatGPT referrers counts a fraction of what happened.

Treat mentions as contributing to pipeline. Claiming they caused it won't survive your own analytics team, and you don't need the claim to justify the work. There are ways to make the picture clearer, which we break down in more depth in this blog post: Does AI visibility actually drive conversions.

What we've seen doing this at volume

You're running an analysis with a sample of thirty. Some of this we can just hand you. The mistakes worth avoiding are collected here.

1. 1

   Corrections convert better than additions, because a publisher fixing an error has an editorial reason to act.

2. 2

   Around 90% of third-party mentions come from listicles, comparisons and reviews. Guides and explainers are where effort goes to die.

3. 3

   Below four live mentions, gains tend not to hold.