We Pitched 30 Journalists AI-Era Angles: What Got Replies

Over six weeks this year we ran an outreach experiment we should have run three years ago. We picked thirty journalists - by hand, one at a time, each one verified as actively covering the specific subject we intended to pitch - and we sent each of them exactly one pitch. Not a sequence. Not a blast. One email, written for that person, plus at most one follow-up.

What made it an experiment rather than a campaign was that we did not send the same kind of pitch to everyone. We built five distinct angle families, assigned six journalists to each, and then measured which ones came back. The angles were the variable. The list quality, the sender, the writing standard and the six-week window were held as constant as we could make them.

Ten of the thirty replied. Four placements ran. But the interesting number is not the thirty-three percent overall reply rate, which is mostly a compliment to the list rather than to us. The interesting number is the spread. One angle family replied at sixty-seven percent. Another replied at zero - six for six, ignored, from journalists who were unambiguously the right people to receive it. Same sender, same weeks, same standard of personalisation, same beats. The only thing that changed was what we were offering, and it changed everything.

This piece is the full experiment: what we mean by an AI-era angle and why that framing matters now, the method in enough detail that you can copy it, the angle-by-angle results including the failures, what correlated with replies beyond the angle itself, the pitch anatomy that actually worked, the citation payoff we were really chasing, and the honest limits of a study this size.

Why We Ran It: Links Were Never The Whole Prize

The reason to re-run an old question is that the answer changed underneath it.

For a decade, outreach to journalists was scored on one axis: did it produce a link. That framing was never quite right, but it was close enough to be useful, and it survived because links were measurable and everything else was not. We wrote about the collapse of the version of this discipline built purely on link acquisition in How Digital PR Replaced Link Building in 2026, and the foundational mechanics in What Is Digital PR.

What has changed is that earned coverage now does a second job, and for many brands the second job has become the more valuable one. When an AI assistant answers a category question, it is deciding which brands to name. That decision leans hard on corroboration - on whether sources other than the brand itself say the brand exists, does this thing, and is credible at it. We measured how badly most brands fail this test in our audit of fifty D2C brands, where thirty-eight never surfaced once in any AI answer for their own category questions. A recurring cause was the absence of any third-party mention the engine could lean on.

So a placement in a real publication is no longer just a link. It is a corroborating source that a retrieval system can find, quote and attribute. That reframe changes what you should be pitching, because the angles that produce citable coverage are not the same angles that produce a friendly mention. That is the hypothesis we set out to test.

What We Mean By An "AI-Era Angle"

The phrase risks sounding like jargon, so here is the working definition we used, which is deliberately narrow.

An AI-era angle is a story angle constructed to satisfy two readers at once. The first reader is the journalist, who decides whether a story exists. The second reader is the answer engine that will later summarise whatever the journalist writes. The angle has to survive both.

In practice that collapses to one test. Can a stranger lift a single sentence out of the pitch, attribute it to a named source, and have that sentence stand up on its own? If yes, you have an angle. Journalists want it because it gives them something concrete to attribute without doing your thinking for you. Answer engines surface it because a self-contained, sourced, checkable statement is precisely what retrieval systems extract - the same property we exploit deliberately when structuring on-page content for answer engine optimisation.

If the sentence only means anything with your brand's context wrapped around it - "the company announced a strategic partnership to accelerate its category leadership" - then you do not have an angle. You have an announcement wearing an angle's clothes. That distinction ended up being the single biggest predictor of whether anyone wrote back.

The Method

We wanted this to be reproducible, so we constrained it more than a normal campaign would be.

The list. Thirty journalists, built one at a time over roughly two weeks. The qualification rule was strict: we had to be able to name a specific recent article by that journalist that made them the right recipient for this specific angle. If we could not name the article, the name did not go on the list. No database exports, no "marketing editor" job-title filtering, no scraped generic desks. That rule alone eliminated more than half the names we initially considered.

The assignment. Six journalists per angle family, five families. We assigned by fit rather than randomly, which is a real limitation and one we come back to at the end - a journalist who covers research studies was more likely to land in the data group. In exchange we got a realistic test, because in the field you would never send a data angle to someone who has never covered a study.

The send. One email each, individually written, no template merge fields. One follow-up after five working days if there was no reply, in the same thread, three sentences maximum. Nothing after that. Sends were spread across mid-morning on Tuesdays, Wednesdays and Thursdays in the journalist's local time.

The logging. For every pitch we recorded the angle family, the subject line and its word count, the body word count, the send day and hour, whether a reply arrived, what kind of reply it was, and whether a placement ran within eight weeks. We also logged declines separately by reason, which turned out to matter more than we expected.

The Results, Angle By Angle

Here is the whole experiment in one table.

Three things sit underneath that table, and each one is more useful than the headline.

The gradient is monotonic, and it tracks one variable. Read the angles top to bottom and what decreases is not quality of writing or strength of relationship. It is how much finished work the pitch handed over. The data angle handed the journalist a fact they could publish. The contrarian angle handed them a fact plus an argument they had to weigh. The timely angle handed them a quote that needed a story to sit inside. The expert offer handed them a person and asked them to invent the story. The announcement handed them a company update and asked them to find the reader interest. Reply rate fell in exact proportion to how much of the work remained.

Placements concentrated even harder than replies did. Replies ran 67 / 50 / 33 / 17 / 0. Placements ran 2 / 1 / 1 / 0 / 0. Half of everything that ran came from a single angle family that was one fifth of the send volume. If you are optimising a small outreach budget, the practical implication is uncomfortable but clear: it is better to spend three weeks producing one genuinely original number than to spend those weeks writing thirty good emails about things you already have.

The zero was not a list failure. This is the result we most wanted to be wrong. The six announcement pitches went to journalists who cover exactly that space, were personalised to a specific recent article of theirs, were sent at good times, and were as well written as everything else. All six were ignored, and none of the follow-ups moved them. When the only variable that changed produced a total wipeout, the variable is the finding.

Why The Data Angle Won

It is tempting to conclude that journalists like data. That is true but shallow, and it leads people to bolt a survey onto a press release and wonder why it did not work.

The mechanism is more specific. A journalist writing a story needs at least one attributable statement that they did not have to take on faith. Sourcing that statement is genuinely expensive - it means finding a study, checking whether the method holds, or getting someone credible on a call. A pitch that arrives with a number, a stated method, and a named person willing to be quoted about it removes that cost entirely.

What mattered was that the method came with the number. In all four data-angle replies, the journalist asked some version of the same question: how did you get that. Two of them asked for the underlying dataset. In one case the reply was essentially a verification exchange, and the story ran only after we had answered it. This is why the "study" that is really a marketing survey of two hundred self-selected newsletter subscribers tends to die on contact - the number arrives, the method does not survive the follow-up question, and the journalist quietly moves on.

The number does not have to be large or expensive to produce. Our own strongest-performing angle came from an audit we had already run for our own purposes and written up in a blog post. The production cost was close to zero because the work already existed. What we had never done before was package it as something a journalist could use, which is a completely different exercise from packaging it as something a reader could enjoy.

Why The Announcement Angle Died

The six-for-six wipeout deserves more than a shrug, because announcement pitching is still the default mode for most brands and most agencies.

An announcement asks the journalist to answer the question "why would my reader care about this company's news," and it asks them to answer it themselves, from scratch, with no help. That is real work. It is work they have no obligation to do, and they are being asked to do it by a stranger, alongside a queue of other emails that morning - several of which arrived with the work already finished.

There is also a structural problem specific to the AI-visibility case. Announcement coverage, on the rare occasions it runs, produces text that is intrinsically hard to cite. "Company X announced Y" is a statement about a company at a moment in time. It is not a claim about the world that a retrieval system can usefully extract when someone asks a category question six months later. So even when the announcement pitch works, it tends to produce the least durable kind of coverage - which is exactly the wrong trade if your goal is the sort of persistent corroboration that gets a brand named in generated answers.

None of this makes company news worthless. It makes it a supporting detail rather than a reason. In two of our four placements the client's actual news appeared inside the article - as context for who produced the data, one line, near the bottom. That is where announcements earn their place: attached to a story that exists for another reason.

What Else Correlated With Replies

The angle explained most of the variance. But we logged enough to see three secondary patterns, and all three are cheap to act on.

Subject lines that read like headlines beat subject lines that read like introductions. The replied set averaged seven words. The ignored set averaged thirteen. The pattern behind the count is that short subject lines were forced to state the finding, while long ones had room to describe the email - "Sharing some research that might interest you for an upcoming piece" describes an email; "38 of 50 D2C brands invisible in AI search" states a finding.

Offering the raw data unconditionally, in the first email, tracked closely with replies. Nine of the ten replied pitches did this. Two of the twenty ignored ones did. We cannot separate this cleanly from the angle effect, since data angles are the ones with data to offer - but the direction is consistent with everything else in the study, and there is no cost to doing it.

Follow-ups recovered almost nothing. Of the ten replies, eight came to the first email and two to the follow-up, and both of those were within a day of it. Nothing in the study justifies a third touch. If two well-aimed emails have produced silence, the honest read is that the angle was wrong for that person, and the fix is a better angle rather than more persistence. This is the opposite of the assumption baked into most outreach tooling, which is designed around sequences and treats reply rate as a function of touch count.

The Replies That Said No

Four of the ten replies were declines. We nearly logged them as failures, and that would have been a mistake.

Three of the four explained why, unprompted. One had run something adjacent two weeks earlier and did not want to repeat the beat. One held a higher bar for sample size than our study met and said so directly, which was fair. One wanted a national framing where ours was regional. That is three specific, free pieces of editorial calibration from people whose job is judging whether stories are worth telling. No tool sells that.

We now track declines-with-reason as a distinct outcome rather than folding them into a failure bucket, because they behave differently over time. A journalist who declined with a reason read the email, thought about it, and spent their own time telling us why. Across the campaigns we have run since, those names have converted on later pitches at a noticeably higher rate than journalists who never replied at all. Counting them as failures throws away the most qualified segment of the list.

The correct response to a reasoned decline is short: thank them, name the specific thing you will fix, stop. No argument, no re-pitch of the same story with a new subject line. The relationship is the asset, and it is worth more than this quarter's placement.

What The Coverage Did For AI Visibility

This is the part we actually ran the experiment to see.

Four placements is not many. But we were not measuring referral traffic, which was modest and predictable. We were watching whether the underlying claim would start showing up in AI-generated answers, cited to the publications rather than to us - the corroboration mechanic we described earlier.

It did, on the engines that retrieve live sources fastest. Within weeks of the placements going live, questions in the neighbourhood of the topic began returning answers that carried the statistic, attributed to the publications that ran it, with our brand named as the origin of the data. The live-retrieval engines moved first, which is consistent with what we see generally and with the structural differences we set out in how to rank on Perplexity and how to rank on ChatGPT. The engines that lean more on training data are a slower and less observable arc.

Two things are worth separating here. The first is that this is a durable asset in a way a link alone is not - the sentence keeps being quotable for as long as it stays true, and it gets quoted from a source that is not us, which is the entire point. The second is that it only works because the coverage contained an extractable claim. Coverage that says a company launched a product produces nothing an engine can lift into a category answer. This is the practical reason we now argue for building the citable artefact first and the outreach second, the same logic behind restructuring content around the questions LLMs actually get asked.

If you are deciding whether this channel is worth the investment for a particular brand, the framework we use for that call is in which clients are worth a GEO strategy, and the wider context on how fast the underlying search behaviour is moving sits in our AI search statistics reference.

The Pitch Anatomy That Worked

Every replied pitch had the same four blocks. This is not a template to copy verbatim - the specifics are what make it work - but the structure held across all ten.

Block one: the subject line, under eight words, stating the finding. Not describing the email. The test is whether it would work as a headline on the story you want them to write.

Block two: one sentence naming the finding, with the number in it. No greeting beyond their name, no throat-clearing, no explanation of who you are. The first line the journalist reads should be the reason to keep reading.

Block three: two or three sentences of method. What the sample was, how it was measured, over what period, and what the honest limitation is. Volunteering the limitation is not weakness. In two of our four data-angle replies the journalist specifically noted that we had flagged our own constraint, and it read as credibility rather than as a caveat.

Block four: the offer and the door. The raw data available unconditionally, a named person available to answer questions on the record, and one line - one - saying who you are. Then stop. No "let me know if this is of interest," no calendar link, no attachment.

What is deliberately absent matters as much as what is present. No company boilerplate. No founder biography. No paragraph explaining what your agency does. No adjectives applied to your own work - if the finding needs to be described as fascinating, it is not.

Running This Yourself In 30 Days

If you want to reproduce the experiment rather than just read about it, this is the sequence, and it fits inside a month.

1. Days 1-5: find or make the citable artefact. Look first at work you have already done - an audit, a dataset, an internal benchmark, a pattern across your client base. Most agencies and most brands are sitting on at least one number nobody outside has seen. Making something new is the fallback, not the starting point.
2. Days 3-7: write the method down honestly. Sample size, how it was collected, over what period, what it does not prove. If the method cannot survive being written down plainly, fix the study before you pitch it.
3. Days 5-12: build the list, one name at a time. Thirty maximum. The rule is absolute: if you cannot name the specific recent article that makes this person right for this angle, they do not go on the list.
4. Days 10-14: assign angles deliberately. If you are testing, split the list into at least two angle families so you have a comparison. If you are not testing, put everything behind the data angle.
5. Days 12-20: write and send individually. Four blocks. Under a hundred and fifty words. Mid-morning, Tuesday to Thursday, their local time. Log everything as you go - angle, subject word count, body word count, send time.
6. Days 17-25: one follow-up, then stop. Same thread, three sentences, five working days after the original. There is no third touch.
7. Days 20-30: answer verification questions fast, and log every decline reason. The verification exchange is where data-angle placements are actually won or lost. Treat a reasoned decline as an outcome worth recording, not a failure to forget.

Then read the results by angle rather than in aggregate. The aggregate number will tell you how good your list was. The per-angle spread will tell you what to pitch next quarter, which is the only output of this exercise that compounds.

The Limits Of This Study

Thirty pitches is a small sample, and I would rather say what is wrong with it than let you assume more than it supports.

The assignment was not random. We matched angles to journalists by fit, so the data angle went disproportionately to journalists who have a track record of covering studies. That inflates its measured performance relative to a randomised design. We accepted this because the randomised version would test something nobody would ever do in practice, but it means the sixty-seven percent is a ceiling rather than an expectation.

Six pitches per cell is thin. A single additional reply in any group moves that group's rate by around seventeen points. None of the individual rates should be treated as precise. What survives the sample-size problem is the shape - a clean monotonic gradient across five groups, with a total wipeout at one end, is not the kind of pattern that appears from noise alone.

The sender was not neutral. These pitches came from an agency with an existing footprint and a founder with a public track record, which almost certainly helped at the margin. A brand starting from nothing should expect lower rates across the board, though there is nothing in the data to suggest the ordering of the angles would change.

And it was one six-week window, in one set of beats. News cycles vary, and a heavy week for a beat suppresses everything unrelated to it.

We are repeating the experiment with a larger list and a stricter design. When we have enough cells to say something firmer than "the shape looks like this," we will publish that too - including whatever contradicts what is above.

What We Actually Changed

Three things changed in how we run outreach, and none of them is about writing better emails.

We now refuse to start an outreach campaign without a citable artefact. If the brand does not have a number, a dataset or a defensible framework, the first phase of the engagement is producing one - not building a media list. That reorders the work and it lengthens the front end, and it has been the single biggest improvement in outcomes.

We cut list sizes hard and moved the saved hours into research. Thirty deeply researched names now beat what three hundred exported ones used to produce, on every metric we track, and it protects the sender reputation that makes the next campaign possible.

And we stopped scoring campaigns on links alone. A placement now gets assessed on whether it contains an extractable, attributable claim - because that is the property that determines whether the coverage keeps working after the news cycle ends. The link is the short-term return. The citation is the compounding one, and the two do not always come from the same story.

If you want the version of this run properly against your own category, that is what our digital PR service is for, and the link building and AI SEO programmes are where the resulting authority gets put to work. If you would rather just get the citable artefact built first, that is a reasonable place to start too - tell us what data you are sitting on and we will tell you honestly whether it is pitchable.