Self-referencing

Evidence

The self-reference effect comes from memory research. Rogers, Kuiper and Kirker (1977) had 32 students judge 40 trait adjectives in four ways — letter size, rhyme, synonymity and the question “Describes you?” — then tested, without warning, which words had stuck. From the self-reference condition participants recalled a mean of 2.84 out of 10 words, from the semantic condition 1.33, from the rhyme condition 0.68, from the structural condition 0.34.

Two things about that paper deserve note. The samples are tiny: 32 people in the first experiment, 27 in the second. And in the second the rest of the ranking collapsed — structural, phonemic and semantic came out level at 0.19 to 0.20, with only self-reference standing apart at 0.32. The authors note in a footnote that their group procedure typically fails to replicate the classic depth hierarchy.

Symons and Johnson (1997) pooled 129 effect sizes. The mean effect is d = 0.50 (95 percent interval 0.45 to 0.54) — a moderate effect, not a large one. How large it comes out depends mostly on what it is compared against:

The more the comparison task resembles self-reference, the less remains. The spread is considerable too: to reach homogeneity the authors had to drop 34 of the 129 effect sizes — 26 percent.

For advertising the bridge is Burnkrant and Unnava (1995) — the only advertising study here. What they manipulated was the copy itself: the high self-referencing version read “you know from your own experience that calculator technology has not improved much”, the comparison version “one knows from experience that …”. Exactly the contrast at issue in forms of address.

The tipping point is not argument strength. It is whether the copy does anything besides the address that already prompts thinking. Both experiments used strong arguments throughout.

The authors’ explanation: self-referencing deepens processing, and with strong arguments that helps at first. Add a second amplifier and processing tips over into tedium and counterargument. In the questions condition, readers’ net positive thoughts fell from 3.71 to 1.50.

What reliably rises is retention. Recall of the ad’s content went from 3.64 to 5.53 (p < .001) — the same finding as in Rogers and in Symons and Johnson, and again only about memory.

Two widespread accounts do not hold. The first: that self-referencing hurts with weak arguments. Weak arguments do not appear in this paper at all. That finding belongs to another paper by the same authors from 1989, and it says something different — there, high self-referencing made strong arguments superior, while under low self-referencing strong and weak arguments did not differ.

The second concerns four conditions under which the address is supposed to work. Two of them are not in the paper that way: the volume of address was never varied, and argument strength belongs to the 1989 paper. The other two hold — low prior involvement and no competing stimuli.

Packard, Moore and McFerran (2018) are the only work here with field data — and their finding goes against direct address. Across six studies they tested which pronoun works in real customer contacts: “I” (the agent), “we” (the firm) or “you” (the customer).

Only the “I” works. Across 1,277 email interactions at an online retailer, linked to purchases over the following 90 days, agent use of “I” predicts higher revenue (β = 1,245.76, p < .001); “we” (β = −201.75, p = .18) and “you” (β = 108.54, p = .33) do not. The β says how strongly each word choice goes together with later revenue; the p-value beside it says how well that link is secured against chance — from .05 it counts as meaningful.

Addressing the customer is the paper’s null finding. The authors tried four times to overturn it; a Bayesian analysis across all the tests puts “no effect” over a hundred times more likely than “there is one”. Ordinarily a study can only fail to show an effect; this method supports the conclusion that the data actively argue for its absence. This concerns service contacts, not ad copy — but it is the best field evidence on direct address there is, and it points the same way as the boundary conditions in Symons and Johnson.

Limits

Memory is not persuasion. The well-supported part of the effect concerns what sticks. That better retention leads to more agreement or more purchases is a second step the classic studies do not take.

With nouns it disappears. Against reference to another person the effect is d = 0.53 for trait adjectives and d = 0.03 for nouns (interval −0.08 to 0.13) — that is, zero. Almost all the studies use trait adjectives; products, services and offers are nouns.

Against an evaluative question it disappears too. Where the comparison task is “Is this word desirable?”, the effect is d = 0.16 (interval −0.06 to 0.37) and indistinguishable from zero. Ad copy asks desirability questions almost throughout.

It depends on being incidental. Against reference to other people the effect is d = 0.02 (−0.14 to 0.18) when participants expect a memory test, and 0.41 when they do not. An expected test is not the same as an expected sales pitch — but the finding shows the effect vanishes once processing becomes deliberate.

Address can also hurt. In Packard and colleagues it turns negative as soon as the customer is not the recipient of the action (“your order is being retrieved”) but is made the actor (“if you have your username, you can look into the account”). Satisfaction and purchase intentions fell from 4.88 to 4.41.

The advertising study is small and student-based. 61 and 75 students, a single product — a calculator from an invented brand — in a lab and a classroom. The authors themselves state that the copy strategy would need testing on samples more representative of an advertiser’s target market before being used.

Attitude toward the ad was untouched. In the first experiment self-referencing did not change it (F < 1.4). What changed was attitude toward the product alone.

With recognition it shrinks. Against semantic processing it is d = 0.69 for free recall and 0.21 for recognition. Online, recognition is the normal case: the offer is still there on the screen.

The effect amplifies in both directions. If self-referencing deepens processing, a weak argument does not become more convincing but more conspicuous. For copy that means: direct address is no substitute for a good argument, only an amplifier for the one you have.

Too much closeness backfires. An address claiming more familiarity than the relationship supports can provoke resistance; see reactance. That link is not verified.

Two different things are called “self-referencing”. In the marketing literature “self-referential advertising” often means something else: advertising that points at itself and comments on its own making. That has nothing to do with the memory effect. Search for the term and you find both mixed together — occasionally within one paper.

This does not cover the German du/Sie distinction. The studies come from English, where “you” carries no polite form. There self-reference is tied to the pronoun; in German it is not — “Sie” is exactly as self-referential as “du”.

The evidence concerns retention, not buying. What was measured is almost entirely how well words and ad content are remembered — not whether anyone buys anything.

The effect depends heavily on the condition. Symons and Johnson (1997) pool 129 effect sizes and find a mean of d = 0.50 — but in several conditions the effect is indistinguishable from zero: 0.03 with nouns, 0.16 against a desirability question, 0.02 with an announced memory test. It is therefore reliable only under particular circumstances, and which ones decide the order of magnitude.